Two popular methods for soft clustering are fuzzy clustering and possibilistic clustering. Possibilistic c-means (PCM) clustering is less sensitive to noise and outliers in data than fuzzy c-means clustering (FCM), but PCM may find non-distinct (duplicate) clusters. Both PCM and FCM can also be used to find single clusters, but FCM yields trivial clusters (the complete set with all memberships equal to one), whereas PCM can find meaningful single clusters, which is called possibilistic one-mean (P1M). P1M is also used for multiple clusters by sequentially finding one cluster after the other, the so-called sequential possibilistic one-means (SP1M), which can avoid non-distinct clusters, as an alternative to conventional PCM. P1M clustering can be done by alternating optimization (AO) of the P1M objective function, but P1M AO may converge to local minima. To the best of our knowledge this is the first paper that examines local minima of the P1M objective function. Moreover, we propose a particle swarm optimization (PSO) algorithm to avoid such local minima. Experiments with synthetic and real-world data indicate that PSO finds the global minimum of the P1M objective function significantly more often than AO, and also on average leads to lower values of the P1M objective function, which yields a higher probability of obtaining better clustering results, at the cost of longer run time, but still feasible even for larger data sets.

This paper proposes a feature-weighted Minkowski distance-based fuzzy k-nearest neighbor regression method, called FWMD-FKNNreg. The key novelty of FWMD-FKNNreg is its integration of feature weighting directly into the Minkowski distance computation in FKNN regression, enabling adaptive control of each feature’s contribution. To achieve this, we introduce four new feature weighting schemes based on relevance, redundancy, and dependency, that are developed using fuzzy mutual information and Łukasiewicz similarity measure. An empirical evaluation identifies the most effective weighting strategy for FKNN regression, with relevance-based weighting outperforming others. Based on this finding, relevance-weighted FWMD-FKNNreg is evaluated on eight real-world datasets and compared with six competitive regression methods, including KNNreg, FKNNreg, Md-FKNNreg, support vector regression (SVR), LASSO, and multiple linear regression (MLR). Experimental results show that FWMD-FKNNreg achieves higher R2 and lower RMSE values in most cases, indicating that the proposed method offers a robust, simple, and effective alternative for regression problems.

This paper presents a novel hierarchical clustering methodology grounded in a refined formulation of a sequential preference stability measure. This measure captures the chance that two agents preserve identical opinions on an alternative across consecutive time points, while remaining well-defined in the presence of incomplete preference information. We formalize the measure, and derive a decomposition that separates within group and cross group stability contributions for arbitrary agent partitions. Leveraging these theoretical results, we introduce the Stability Hierarchical Clustering Algorithm. An efficient incremental update strategy substantially reduces computational overhead. The methodology is validated by means of two case studies: a synthetic temporal preference profile designed to illustrate its discriminative capability, and an empirical dataset concerning life-support decisions of cancer patients, which demonstrates its practical applicability.

Hard Possibilistic C-Means (HPCM) is a prototype-based clustering method that replaces forced assignments with binary typicalities, allowing a point to be assigned to multiple clusters (overlap) or to none (rejection). This makes HPCM attractive in settings with ambiguity and background noise, but it also complicates evaluation and raises practical questions about when HPCM behaves differently from standard Hard C-Means (HCM). In this paper we present an empirical study of HPCM through controlled synthetic experiments and small real-world benchmarks. We visualize how HPCM differs from HCM in overlap and noise regimes, and we analyze the sensitivity of HPCM to the scale parameter that defines its acceptance thresholds, showing how uniform background noise can inflate it and change the effective rejection/overlap behaviour. Since HPCM does not output a classical partition, we evaluate it using set-valued/reject-style metrics that capture the accuracy--coverage and accuracy--set-size tradeoffs. Overall, the results highlight both the intended advantages of HPCM (explicit ambiguity and noise filtering) and the conditions under which its behaviour can degenerate due to scale estimation.

Detection systems for stress monitoring in airport settings must be precise, comprehensible, and deployable under practical limitations. A comparative benchmarking analysis of wearable physiological sensing-based stress detection techniques is presented in this paper, which contrasts cutting-edge deep learning and computer vision techniques with an interpretable fuzzy inference framework based on Design Science Research. The analysis takes into account typical wearable signals, such as electrocardiography (ECG), electrodermal activity (EDA), and heart rate variability (HRV), and assesses the consequences for real-time monitoring of both software and hardware deployment. Several benchmark datasets are used to test methods based on deployment-oriented characteristics, including robustness, interpretability, data requirements, computing cost, privacy, and sustainability. The results demonstrate that the fuzzy inference method offers transparent, stable, and privacy-preserving stress estimate with low computing overhead, while deep learning and vision-based models reach high accuracy in controlled circumstances. For ethical and sustainable airport applications, these results encourage the use of interpretable wearable-based stress monitoring.

Abstract— Accurate short-term energy demand forecasting is essential for the stability and efficiency of modern energy smart grids, especially with the growing integration of renewable energy sources, which introduce significant uncertainty. Traditional machine learning models, such as neural networks and deep learning, are "black boxes" that deliver accurate predictions but lack transparency, limiting operator trust and actionable insights. To help users better understand the prediction model and increase their confidence in its outputs, it is necessary to employ eXplainable Artificial Intelligence (XAI), whose models and predicted results can be easily understood, analyzed, and improved by business users. In this paper, we present an Interval Type-2 Fuzzy Logic System (IT2 FLS)-based XAI system, designed to manage noise, non-stationarity, and uncertainty in hourly demand data. We utilize SHapley Additive Explanations (SHAP) and a Random Forest (RF) to enhance model interpretability and prediction reliability by reducing the number of inputs and rules in the IT2 FLS. In our approach, SHAP is used to rank features and combine their importance with RF-based significance for optimal feature selection and rule pruning in the IT2 FLS. This method reduces the root-mean-squared error (RMSE) on testing data by 55%, from 5.94 kilowatt-hours (with a full rule base of 2,781 rules) to 2.74 kilowatt-hours, using only three inputs and 16 rules—a 250-fold reduction in the number of rules. In comparison, the Type-1 FLS showed lower performance with an RMSE of 5.376, while the neural network achieved a lower RMSE of 1.410 but lacked interpretability. The proposed IT2 FLS-based XAI system thus offers a practical balance between accuracy and explainability, enabling real-time smart grid applications. This approach will be further applied within a smart grid to facilitate adaptive decision-making and optimization.

The convergence of neuroscience and artificial intelligence has led to the development of "living biocomputers," where biological neural networks are cultured on Micro-Electrode Array (MEA) chips. This creates hybrid biocomputing platforms that blur the boundaries between biology and technology. A critical challenge in this domain is the accurate and interpretable detection of neural spikes from MEA recordings. While black-box models like deep neural networks have achieved good accuracy, their lack of transparency hinders scientific inquiry. Revealing spatial-temporal patterns across electrode arrays can inform electrode placement strategies, stimulation protocols, and our understanding of how biological neural networks process information. This approach could bridge the gap between understanding the organization of neurons, their activity and the computing task in hand and ultimately allow living neural biocomputers to have a solution to the current “black-box” biocomputing approaches. We adopt a fuzzy logic framework because it can handle the biological uncertainty and signal variability inherent in MEA recordings, enabling more faithful spike discrimination than rigid threshold-based or crisp machine learning approaches. This paper presents a fuzzy rule-based classifier that combines high performance with human-readable interpretability for spike detection in neural computing. Our approach integrates ANNIGMA (Artificial Neural Network for Input Gain and Measurement Analysis)-based feature selection to identify the most salient spike characteristics and a genetic algorithm to optimize the fuzzy rule base. We validated our model using multi-chip pooled validation protocol, achieving a peak F1 score on testing data of 97.74% on a two-chip experiment and 96.12% on a six-chip experiment. The resulting fuzzy rules are linguistically interpretable, allowing researchers to understand the underlying logic of spike detection. This work can be a step forward towards building trustworthy neural biocomputers with direct applications to drug discovery, disease modeling, and understanding the fundamentals of neural computation.

We introduce an interpretable clinical classification framework that couples Adaptive Neuro-Fuzzy Inference Systems (ANFIS) with Fuzzy Random Forests (FRF) to achieve high performance while maintaining transparent decision logic. In the proposed design, ANFIS learns data-driven Gaussian membership functions and optimizes their mean and sigma parameters, which are then used to induce a robust FRF ensemble for decision-making. The end-to-end pipeline follows a leakage-aware methodology: (i) stratified train–test splitting with skewness-aware imputation computed strictly from the training set; (ii) Z-score normalization (zero mean, unit variance) fitted on training data only; (iii) mutual information–based feature selection via SelectKBest to obtain a compact, discriminative subset; (iv) SMOTE class balancing applied exclusively to the training set; (v) ANFIS-based optimization of Gaussian membership parameters; and (vi) FRF construction using ANFIS-optimized fuzzy partitions to produce interpretable predictions. Experiments on two UCI benchmarks (Cleveland Heart Disease and Wisconsin Diagnostic Breast Cancer) show that the ANFIS–FRF framework achieves AUC values of approximately 0.95 and 0.99, respectively, matching or exceeding strong baselines such as Random Forest and XGBoost. Importantly, the model yields a human-readable rule set, whose interpretability is further examined through an LLM-assisted evaluation. By decoupling fuzzy feature learning (ANFIS) from decision logic induction (FRF) and embedding mutual information-driven feature selection within rigorous preprocessing, the proposed approach provides an accurate, robust, and transparently explainable solution for clinical decision support.

Minimum cost consensus (MCC) models are widely used in group decision-making to achieve agreement with minimal opinion adjustment costs. Although recent studies have incorporated fairness into MCC frameworks, most existing approaches rely on aggregate social welfare or inequality-based axioms and remain insensitive to subgroup structures induced by protected attributes. As a result, consensus outcomes may still systematically favor certain subgroups despite being cost-optimal or globally fair. To address this limitation, this paper proposes a fairness-aware MCC framework that explicitly accounts for subgroup-level fairness. Fairness is defined through deviation-based measures that evaluate the parity between the final consensus outcome and the opinions of subgroups formed by protected attributes. Fairness is incorporated into MCC models via constraint-based formulations and a learning-inspired objective that interpolates between utilitarian and Rawlsian principles. The proposed models preserve convexity and computational tractability, while enabling a flexible balance between efficiency, consensus, and subgroup fairness in heterogeneous group decision-making environments.

To address the ethical and societal risks of artificial intelligence, responsible artificial intelligence that ensures not only high accuracy but also transparency and fairness is increasingly demanded. Inherently interpretable models are valuable in scenarios where both transparency and fairness are important, as the model’s internal mechanisms can be inspected to support bias diagnosis and mitigation. Fuzzy systems are representative inherently interpretable models that enable flexible decisions considering real-world uncertainties. Multi-objective fuzzy genetics-based machine learning generates a diverse set of fuzzy classifiers that consider trade-offs among multiple objectives using an evolutionary multi-objective optimization algorithm. However, previous studies on fairness in fuzzy classifiers have focused exclusively on group fairness and have not considered individual fairness. In this study, we propose a multi-objective fuzzy genetics-based machine learning that simultaneously optimizes accuracy and individual fairness, and investigate how incorporating individual fairness as an objective affects the resulting set of fuzzy classifiers. Experimental results show the effectiveness of this approach and provide empirical insights into the relationships among accuracy, group fairness, and individual fairness, contributing to fair AI design. Source code is available at https://github.com/TakeruKonishi/MoFGBML Fairness.

Overlapping community detection in complex networks remains a fundamental unresolved problem due to the fuzzy nature of node memberships and the computational bottleneck imposed by the O(n^2) complexity of conventional density-peak algorithms. Real-world networks are characterized by continuous membership distributions in which nodes belong to multiple functional groups. However, existing methods often compromise this structural realism by imposing discrete assignments or are computationally infeasible for networks at the million-node scale. In this paper, we propose DPC-HOPE, a framework that uses Density Peak Clustering (DPC) using a Structural Fuzzy Membership Propagation lens. We replace Euclidean distance with a High-Order Proximity Embedding (HOPE) and then introduce a Fuzzy Directed Acyclic Graph (F-DAG) propagation mechanism to model the gradient of membership flow from community peaks to the boundaries. This enables an explicit, non-heuristic determination of overlapping memberships with a computational complexity of O(m + n log n). Experimental results on large-scale LFR benchmarks and real-world networks demonstrate that DPC-HOPE consistently outperforms existing community detection algorithms in both accuracy and community quality. These results suggest that explicitly capturing topological fuzziness is essential to achieving both scalability and accuracy in large-scale network analysis.

This study is inspired by quantum algorithms and proposes the Quantum-inspired Search Algorithm (QiSA), an entropy-driven quantum-inspired metaheuristic for parameter-free combinatorial optimization. Guided by Occam's razor, QiSA simplifies the search process and reduces parameter dependence as much as possible in this preliminary study. It uses entropy as the core signal to unify parameter control, overcoming the traditional dilemma of balancing exploration and exploitation, and maps it to a confidence-like level to define an interpretable closed-loop termination criterion. QiSA also exhibits several observable quantum-like behaviors during the search, reflecting characteristic phenomena associated with quantum algorithms. Experiments on classic combinatorial optimization benchmarks show that QiSA delivers robust, consistent, and competitive performance, with the potential to stand alongside Differential Evolution in combinatorial optimization.

Recently the multi-child differential evolution (MCDE) algorithm, in which each parent competes against more than one child, has been introduced. It has been demonstrated that MCDE achieves a wall-clock time reduction of up to two orders of magnitude for massively parallelized optimization. Results which were based on the classical DE algorithm DE/rand/1/bin showed that the speedup achieved by MCDE scales well with the number K of processors, even when K is orders of magnitude higher than the population size N. In this work the multi-child version of SHADE is compared with MCDE yielding promising results.

Multi-agent path finding (MAPF) is a critical problem in applications such as warehouse automation and autonomous driving. In real-world settings, multiple conflicting objectives, such as minimizing path length and ensuring safety, often must be optimized simultaneously, motivating the study of multi-objective MAPF (MOMAPF). However, existing MOMAPF approaches suffer from severe scalability limitations as the number of agents and the environment size increase. To address this challenge, we propose Two-Stage Evolutionary MOMAPF (TSE-MOMAPF), which optimizes a single agent's path via NSGA-II and refines the path set using an SMS-EMOA-based framework. Experiments on various grid maps show that the proposed method achieves superior scalability and provides broader Pareto front coverage in medium- and large-scale scenarios compared to an existing method. These results indicate that the proposed evolutionary framework effectively balances solution diversity and computational efficiency under strict time constraints.

Reinforcement learning-assisted evolutionary algorithms are attracting increasing attention, but existing methods rely on handcrafted state representations that are often problem-dependent, limited in expressiveness, and lack robustness to variations in population composition and distribution. To address these limitations, a self-supervised state representation learning framework is proposed for reinforcement learning-assisted evolutionary algorithms. This framework follows a two-stage learning paradigm consisting of representation learning and policy learning, where the state encoder is trained in the representation learning stage and subsequently used to generate state inputs for reinforcement learning based policy inference. Specifically, a set-aware attention-based encoder is designed and trained via self-supervised contrastive learning using population objective values, while an auxiliary guided learning strategy with dynamic weighting is employed to stabilize the representation learning process. To evaluate the effectiveness of the proposed framework, it is integrated into representative reinforcement learning-assisted evolutionary algorithms for single- and multi-objective optimization problems to assess its generalization capability, with additional ablation studies analyzing key components. The results demonstrate the effectiveness of the proposed framework in addressing the limitations of handcrafted designs.

Artificial Immune Systems (AIS) have been extensively applied to the single-class anomaly detection problem, in optimization and adaptive control system applications. However, recent works are are finding ways to adapt AISs to address the needs of multi-class problems, at times achieving competitive advantages over traditional machine learning. Further, for higher dimensional problems such as Fake News classification, further extensions are needed. One such extension is Unbounded Recognition regions -- the area of the search space in which an antibody may detect an antigen (data). Exciting hybrid advances are also present in the field, combining AIS with, for example, Island Genetic Algorithms or Quantum computing.

The work herein introduces the AISIGA model - a hybrid AIS and Island Genetic Algorithm, adapted from the MAIM model. Results exploring the extension of unbounded recognition regions, contributes to an improved understanding of this newer AIS mechanism. Preliminary results comparing AISIGA with the state of the art, show promise but a need for further work, with respect to accuracy whilst highlighting the significant efficiency benefits of AISIGA compared to the state of the art.

This paper addresses the inverse problem of reconstructing basement relief from gravity data by comparing two contrasting optimization strategies: a linearized Dual-Simplex (DS) method and a Memetic Algorithm (MA) combining global evolutionary operators with Gauss–Newton local refinement. The DS method solves a Linear Programming (LP) formulation based on $L_1$-norm data misfit and smoothness regularization, whereas the MA couples global exploration (BLX-$\alpha$ crossover, hybrid mutation, and tournament selection) with a weighted local update. Both methods are applied to synthetic and real 2D datasets—the Isolated Graben model and the Poem Bridge profile. In the synthetic case, characterized by smooth geometry and weak nonlinearity, DS achieves higher accuracy, including a 1\% relative error in 98\% of the observations. In contrast, in the more heterogeneous and nonlinear Poem Bridge dataset, the MA yields better cumulative relative data misfit error reduction, achieving a 5\% error in roughly 72\% of the data, compared to 58\% for DS. Overall, deterministic linearized methods prove efficient for well-behaved models, while evolutionary strategies offer greater robustness under complex real-world conditions. The results emphasize the complementary nature of both approaches and indicate that the optimal inversion strategy should be guided by the geological complexity and nonlinearity of the problem.

Wind farm layout optimization involves multiple conflicting objectives and highly irregular search landscapes, making the selection of suitable multi-objective metaheuristics a challenging task. Although meta-learning has been successfully employed to select metaheuristics for single-objective optimization, its application to real-world multi-objective engineering problems remains unexplored. This paper proposes a meta-learning approach for automatically selecting metaheuristics for the Multi-Objective Wind Farm Layout Optimization Problem (MoWFLOP). The proposed approach uses fitness landscape meta-features to predict the performance of multi-objective state-of-the-art metaheuristics under different quality indicators. We evaluate 90 datasets with different sampling configurations from 516 wind farms. The analysis comprised the results of the metaheuristics, the impact of sampling configurations, and the meta-learning performance for metaheuristic selection. Results show that the meta-learning approach surpassed the baseline of selecting the same algorithm for all instances.

Abstraction and reasoning are essential abilities in human intelligence, which help humans understand the world and extend our knowledge to unseen cases. However, abstraction and reasoning are still big challenges for existing artificial intelligence models due to limited training instances and domain-specific concepts. Existing studies usually model the abstraction and reasoning tasks as program synthesis tasks. Genetic programming features in program synthesis. However, genetic programming is hardly applied to abstraction and reasoning tasks. To fill this gap and explore genetic programming's capability for abstraction and reasoning, this paper proposed a typed linear genetic programming method and evaluated its performance on the Abstraction and Reasoning Corpus. The results show that the typed linear genetic programming has a superior performance to many existing methods, implying a good potential to fulfill abstraction and reasoning in artificial intelligence.

Gene Expression Programming Neural Networks (GEPNN) evolve neural network topologies through a fixed-length linear chromosome that decodes into expression trees. This encoding guarantees syntactic validity under genetic operations, but inherently produces acyclic structures that prevent GEPNN from solving temporal problems that require memory. This paper introduces the index terminal, a novel terminal type that references other nodes within the expression tree by their position on the chromosome. When a node references an earlier position, a recurrent connection is formed; when it references a later position, a skip connection is created. This mechanism enables arbitrary cycles in the computational graph while preserving fixed-length encoding and deterministic decoding properties found in Gene Expression Programming (GEP). The approach is validated on temporal XOR benchmarks that require one- and two-step memory. The results obtained confirm that evolution discovers recurrent connections through index terminals that maintain state across timesteps, enabling GEPNN to solve tasks previously inaccessible to the framework. Delayed regression experiments confirm that the mechanism generalizes to continuous-output tasks. This shows that index terminals support temporal sequence processing in both the classification and regression domains. These results establish that the proposed index terminals successfully extend GEPNN to temporal problems while maintaining the algorithmic simplicity that distinguishes GEP from variable-length neuroevolutionary encodings.

The additive Gaussian noise mutation approach is commonly used in state-of-the-art neuro-evolution algorithms to optimize/train the weights of networks with fixed topology. In this paper, a novel modification to this mutation approach is proposed and studied. The proposed modification follows a theoretical analysis of mutations by the additive Gaussian noise. The well-known neuro-evolutionary algorithm of the Uber AI labs, which is based on additive Gaussian noise, is applied here as a reference algorithm. First, we modified the reference algorithm by replacing its mutation mechanism with the proposed mutation technique. Next, a comparison study is reported between the modified algorithm and the reference one. The comparison study is conducted on several well-known continuous control benchmark problems. The theoretical and computational studies indicate that the proposed modifications are not subjected to the constraints of the additive Gaussian noise approach, which appears to provide significantly better results with increasing problem dimensions.

While deepfake speech detectors built on large self-supervised learning (SSL) models achieve high accuracy, employing standard ensemble fusion to further enhance robustness often results in oversized systems with diminishing returns. To address this, we propose an evolutionary multi-objective score fusion framework that jointly minimizes detection error and system complexity. We explore two encodings optimized by NSGA-II: binary-coded detector selection for score averaging and a real-valued scheme that optimizes detector weights for a weighted sum. Experiments on the ASVspoof 5 dataset with 36 SSL-based detectors show that the obtained Pareto fronts outperform simple averaging and logistic regression baselines. The real-valued variant achieves 2.37% EER (0.0684 minDCF) and identifies configurations that match state-of-the-art performance while significantly reducing system complexity, requiring only half the parameters. Our method also provides a diverse set of trade-off solutions, enabling deployment choices that balance accuracy and computational cost.

Reinforcement Learning (RL) enables the training of autonomous agents to accomplish various tasks through environmental feedback. This feedback, formalized as rewards, guides the agent toward optimal strategies for achieving desired objectives. However, a fundamental challenge exists: agents optimize for reward maximization rather than directly solving the intended tasks. The reward functions, typically designed manually, do not necessarily lead to optimal solutions and may fail to capture the true complexity of the problem domain. Even reward functions that appear intuitive to human designers often produce suboptimal or unexpected agent behaviors.

We introduce RevolveRL, a novel approach leveraging Genetic Programming to discover effective reward functions that enhance performance and reduce training time. Our method employs task-specific fitness functions designed around the ultimate goals of the tasks, which guide the evolutionary search process. We evaluate our approach on both classic RL benchmarks and the more complex MuJoCo physics simulation tasks. Our experimental results demonstrate that automatically evolved reward functions consistently yield superior policies and accelerate the training process compared to baseline manually designed rewards.

Multi-objective evolutionary algorithms (MOEAs) are widely used for solving complex optimization problems involving conflicting objectives. While NSGA-II (Non-dominated Sorting Genetic Algorithm II) remains a strong baseline, recent approaches such as Multi-Objective Factored Evolutionary Algorithms (MOFEA) and novelty-driven search aim to improve scalability and exploration in high-dimensional decision spaces. Based on a hypothesis that introducing novelty search into MOEAs should improve quality of returned Pareto sets, we present a systematic empirical comparison of four algorithms: NSGA-II, MOFEA, Novelty-MOEA (Multi-Objective Evolutionary Algorithm), and Novelty-MOFEA (Multi-Objective Factored Evolutionary Algorithms) on the ZDT test suite (excluding ZDT5) and all DTLZ problems. We evaluate their performance using standard quality indicators including Hypervolume (HV), $k$-nearest neighbor distance, Inverted Spacing (SP), and the UD unary quality indicator. Our results reveal that no single algorithm dominates across all problem types. Instead, performance depends strongly on problem characteristics such as modality, separability, and the presence of deceptive local optima. We analyze these trends and provide insights into when factored optimization and novelty mechanisms are most beneficial.

The basic idea of the hyper-heuristic (HH) algorithm is to integrate multiple search algorithms to leverage their strengths for optimization problems. One of the main challenges in HH design is selecting or generating low-level heuristics (LLHs) effectively within a limited number of evaluations. To address this challenge, this paper proposes an effective HH method based on the machine learning process for solving single-objective real-parameter optimization problems (SOPs). During the offline stage, a set of benchmark functions and an effective metaheuristic algorithm (i.e., search economics) will be used to automatically generate a “good” sequence of LLHs and a set of suitable hyperparameters for them. This good sequence can be regarded as a combination of LLH sets, which will construct a “new” metaheuristic algorithm for solving SOPs during the online stage. The experiments are conducted using the CEC 2024 and CEC 2025 benchmarks to evaluate the performance of the proposed method. The results conclusively demonstrate that this approach significantly outperforms the state-of-the-art algorithms.

Applying knowledge transfer across multiple optimization tasks, multitask optimization (MTO) emerges as a promising approach to solving synergistic optimization tasks simultaneously. However, the development of effective knowledge transfer mechanisms in MTO fundamentally relies on aligning elite solution distributions across tasks. This dependency creates a critical bottleneck in few-shot optimization regimes, as restricted evaluation budgets impede the identification of elite solution distributions required for beneficial transfer. This challenge is exacerbated in multiobjective multitask problems, where each optimizer must approximate a continuous Pareto manifold rather than a single optimal point. This paper introduces Iterative Sequential Transfer (IST) to circumvent this bottleneck. We model MTO as a sequence of sequential transfer optimization problems, concentrating evaluations on a single target per iteration. We propose a likelihood-informed task prioritization mechanism to maximize transfer utility by identifying the task most likely ready for knowledge integration. Empirical results on benchmark and real-world problems verify the effectiveness of the proposed method under tight budgets.

On-site delivery and tuning of Large Language Model (LLM) inference face several challenges, like diverse environments, complex tuning parameters, and unpredictable workloads. We propose Perceptor, an end-to-end adaptive tuning engine to overcome these issues. Based on a perceive-adaption loop, Perceptor achieves lifetime-span adaptive performance tuning across various environments. It introduces a multi-stage optimization pipeline decoupling cross-layer key tuning parameters, and pioneers a collaborative strategy that combines pre-deployment (static) and post-deployment (online) optimizations to always keep optimal inference performance. Specifically in optimization pipeline, we have proposed novel adaptive optimization algorithms for parallel strategy, RoCE networks, and Prefill-Decode (PD) workload scheduling. Experiments demonstrate significant performance improvements achieved by Perceptor across various scenarios.

Continuous relaxations are commonly used as an encoding scheme for evolutionary multitasking optimization (EMTO), allowing heterogeneous tasks to be optimized within a unified search space. However, these continuous representations fundamentally reshape the induced fitness landscape, and their impact on optimization behavior and knowledge transfer remains underexplored. In this work, we bridge the gap by analyzing the interaction between the representation-induced fitness landscape and optimizer behavior through fitness landscape analysis. Using piano fingering estimation as a representative case of the subset selection problem, we introduce several encoding variants and investigate how their induced landscapes influence search dynamics and solution quality under EMTO. Our analysis indicates that transfer efficacy is strongly associated with changes in representation-induced landscape characteristics, rather than algorithmic factors alone, highlighting the representation design as the key factor of effective knowledge transfer in multitasking optimization.

In this paper, we extend the scalable multi-agent pathfinding problem to modifiable environments. Instead of only finding collision-free paths, agents collaboratively alter the environment to transform it to a predefined target state. Applications of this problem include smart warehousing, terrain leveling, autonomous mining, and automated construction. We introduce a scalable simulation environment for such tasks and propose an evolutionary optimization approach. Two experiments are conducted: (1) analyzing the effect of the number of agents on task performance and (2) comparing the performance of different crossover operators. We propose a new two-level crossover operator acting on different layers of the gene encoding and compare it with two one-level operators. Results indicate that more agents reduce the task completion time, but fewer agents yield more consistent solutions with better objective values due to the collision constraints. Moreover, the introduced two-level crossover operator achieves significantly better objective values than both one-level operators in most problem instances.

The Electric Vehicle Routing Problem (EVRP) requires the joint optimization of customer routing and energy feasibility under battery capacity constraints. Although many existing EVRP solution strategies implicitly follow a bilevel structure, routing and feasibility handling are commonly implemented within monolithic solvers, making algorithmic components tightly coupled and difficult to substitute or reuse.

This paper introduces an Iterative Instance Refinement Framework (IIRF) that explicitly decouples routing optimization from feasibility enforcement through a problem-instance interface. The framework links an upper-level VRP solver with a lower-level EV feasibility evaluator via an EV-aware cost matrix to iteratively refine the problem instance for steering the routing solver toward energy-feasible solutions. By construction, the framework is both solver-agnostic and evaluator-agnostic, enabling different routing algorithms, feasibility evaluators, and refinement strategies to be integrated without altering the algorithmic core.

Empirical results show that refining problem instances enables a classical VRP solver to produce high-quality and energy-feasible solutions without explicitly handling charging decisions during its search. On two EVRP benchmark suites, E-X and E-F-M-X, IIRF not only delivers the state-of-the-art performance but also advances the best-known solutions to date for multiple problems. These findings validate the feasibility of the modular design and indicate a pathway toward an evolutionary optimization framework that operates at the level of problem instances, rather than exclusively at the level of explicit solution representations.

Optimization of robotic arm trajectories represents a complex computational challenge critical for efficient and safe automation. While traditional metaheuristic approaches have achieved notable success, they often require extensive manual parameter tuning and lack robustness when problem characteristics change. This work introduces a novel approach to robotic trajectory optimization based on the Modular Differential Evolution (MODDE) framework. Instead of designing a single, problem-specific algorithm, we employ a hyperheuristic framework that autonomously manages and configures the components of MODDE to identify optimal trajectories regarding execution time and motion smoothness. We demonstrate that employing this hyperheuristic control results in more robust and adaptive solutions compared to manually tuned metaheuristics, representing a significant advancement in the practical application of evolutionary computation.

Symbolic regression (SR) automatically derives interpretable mathematical expressions from data and has been widely applied in various fields. Many real-world SR tasks are intrinsically similar, such as fault prediction models for different engine types sharing common structures. However, most existing SR methods focus on independent single-task modeling and fail to exploit inter-task relationships to improve performance. To address this, a task-similarity-based multitask genetic programming (TSMTGP) approach is proposed for SR. TSMTGP adopts a novel two-tree individual representation, in which each solution consists of a shared tree across tasks and a task-specific tree tailored to each task, enabling effective knowledge transfer. Correspondingly, two fitness evaluation methods are designed for the shared tree and the two-tree representation. A task similarity measure strategy is developed by using Spearman’s rank correlation coefficient based on the performance of the shared tree on different tasks. Based on this measure, a similarity-aware evolutionary mechanism is introduced, in which the evolutionary process adaptively switched between high-similarity and low-similarity modes to improve knowledge transfer. TSMTGP is consistently evaluated under varying task similarities on both benchmark and real-world problems. TSMTGP consistently outperforms comparative methods, particularly on real-world datasets, demonstrating its effectiveness in dynamic knowledge transfer across different tasks.

In today's technology focused environment, networks receive a constant multitude of requests from various user devices, leading to cases of high latency in the network if data transmissions are mismanaged. If a denial of service (DoS) or distributed denial of service (DDoS) attack is launched on a network with mismanaged data transmissions, the damage would be catastrophic. Thus, appropriate management of transmissions is paramount. Methods for managing communications exist, but they require costly computations and account for select variables such as energy consumption or the amount of data transmitted. Our approach utilizes genetic algorithms to generate network configurations with structures that reduce latency based on different factors: data transmitted over a connection, and energy consumption. Once designed, their resilience to DoS/DDoS attacks is evaluated. The results show the methodology designs networks resistant to attack and that the surviving networks can also be reconfigured to re-establish the fitness to a pre-attacked state.

Surrogate-assisted evolutionary algorithms (SAEAs) alternate expensive evaluations using surrogate models of objective functions, and thus have become a key approach for solving expensive optimization problems. Several efforts have been made to enhance the prediction accuracy of surrogate models. However, this does not necessarily lead to performance improvement of SAEAs, as solutions contributing to the update of the best objective value and model refinement are technically different from each other in nature. Thus, balancing the two aspects is crucial for accelerating the performance of SAEAs. Nevertheless, the resource allocation method to sample solutions for optimization or for model refinement remains unclear. Accordingly, this work provides a novel structure-aware adaptive switching methodology between evolutionary search and model refinement modes, grounded in explainable machine learning. Specifically, SHapley Additive exPlanations (SHAP) are employed to quantify the structural instability of the surrogate model. When the surrogate model is structurally unstable or inaccurate, inspired by active learning, the proposed SAEA framework generates solutions by anisotropically perturbing decision variables according to dimension-wise SHAP instability to refine the model. Otherwise, it switches to the evolutionary search mode to discover better solutions. The effectiveness of the proposed framework over state-of-the-art SAEAs is demonstrated through comparison on benchmark and engineering design problems.

The Electric Vehicle Charging Scheduling Problem (EVCSP) concerns the allocation of limited charging resources to heterogeneous vehicle requests under temporal and capacity constraints. Although advanced optimization models have been proposed for this problem, most assume an offline setting with complete knowledge of future arrivals, limiting their applicability in real charging stations. This paper introduces an online, preemptive EVCSP that preserves a high-fidelity problem formulation, including charger availability and grid capacity constraints, while enforcing a commitment policy under which admission decisions are irrevocable. We propose a hybrid solution approach that combines a genetic algorithm for online admission control and charger assignment with a mathematical optimization model for energy allocation. To meet real-time decision-making requirements, we embed a Gaussian-process surrogate within the evolutionary search to reduce the cost of repeatedly evaluating candidate schedules. Computational experiments across multiple instance sizes and booking modes show that surrogate assistance preserves objective values relative to exact evaluation while yielding statistically significant reductions in computational time, supporting practical online deployment of evolutionary optimization for real-time EV charging station operations.

Artificial neural networks (NNs) often struggle to maintain performance when concept drift, a phenomenon where the underlying data distribution changes over time, occurs. While traditional training approaches based on backpropagation struggle to maintain performance under drift conditions, ensemble learning combined with dynamic multi-modal (DMM) optimization offers a promising alternative. This paper investigates the applicability of multi-quantum swarm optimization (MQSO), a DMM particle swarm optimization (PSO) variant, for training NN ensembles in dynamic environments. MQSO is evaluated against traditional gradient-based methods and standard quantum swarm optimization. MQSO achieves superior performance on lower-dimensional problems with simpler decision boundaries, effectively tracking multiple optima. However, gradient-based methods with retraining outperform MQSO on higher-dimensional, non-linear problems due to computational constraints when using MQSO. The findings further show that retraining mechanisms are essential for maintaining performance under concept drift. It is also clear that DMM PSO is particularly effective for tracking rapidly changing decision boundaries.

This paper proposes a particle swarm optimisation maximum-margin classifier (PSO-MMC) that borrows core principles from support vector machines (SVMs). The task of identifying an optimal linear separating hyperplane is formulated as a constrained optimisation problem defined exclusively over training instances that form Tomek links. For non-linearly separable data, kernel mappings are incorporated in a manner analogous to the SVM kernel formulation. Prior to Tomek link identification, noise is removed from the training data using established filtering techniques. Experimental evaluation is conducted against classical SVMs, classification trees and 𝑘-nearest neighbour (KNN) across multiple benchmark data sets. The results show that the proposed classifier achieves improved classification accuracy and reduced optimisation complexity, particularly under conditions of class overlap and feature redundancy. The findings indicate that restricting optimisation to boundary-critical instances provides an efficient and scalable alternative model for maximum-margin classification.

Most RNA inverse folding methods rely on a single structure predictor, often resulting in sequences that overfit specific algorithmic biases. To address this, we present a predictor-agnostic Evolutionary Algorithm framework that decouples optimization logic from folding engines. On the Eterna100 benchmark, our framework achieves a 68% solve rate using single-tool modes, matching the performance of state-of-the-art reinforcement learning methods. Crucially, to mitigate the risk of model-specific errors, we introduce a consensus optimization strategy that seeks sequences validated by multiple, distinct folding models (e.g., thermodynamic and deep learning). By filtering out predictor-specific artifacts, this approach acts as a structural "safety net," effectively reducing the occurrence of extreme prediction outliers and yielding more robust candidates for downstream experimental testing. Finally, the framework features a modular "plug-and-play" architecture that supports granular, nucleotide-level constraint control, enabling precise customization for diverse design tasks.

The population-based ant colony optimization (P-ACO) algorithm has been proven effective in addressing the dynamic electric vehicle routing problem (DEVRP). In the DEVRP the travel time is subject to dynamic changes representing real-world traffic conditions. P-ACO maintains an archive of solutions that are used to update the pheromone trails. By default, the archive is updated in first in, first out fashion. In this study, we introduce a new update strategy based on the similarities of the solutions stored in the archive to maintain the diversity when updating the pheromone trails. The experimental results on different DEVRP test cases demonstrate the effectiveness of the proposed strategy in comparison with other existing ones.

To overcome the limitations of conventional evolutionary algorithms in dynamic integrated energy systems for steel enterprises, this paper proposes a Conditional Variational Autoencoder (CVAE) guided multi-objective optimization framework. The method minimizes operational costs and external energy reliance by using online-trained CVAE to map dynamic environmental states (e.g., energy loads, price signals) to optimal solution distributions. Upon detecting changes, the model rapidly generates a state-conditioned initial population, providing a warm start for the evolutionary search. Experiments verify that this framework significantly accelerates convergence and improves solution stability and economic performance in dynamic environments.

In the dynamic multimodal optimization problems where multiple optimal solutions change their locations and their number as time passes, enough number of individuals are indispensable to track the dynamically changing solutions. For this issue, many conventional methods take an approach of increasing the number of individuals becasue an optimal number of individuals are not known, but such an approach makes it hard to track the solutions when their speed become fast. To tackel this problem, this paper proposes (i) NCCO (sup+ul) which incorporates the suppression mechanism and an upper-limit mechanism into NMMSO- and CMA-ES-based Continuous Optimizer (NCCO) and (ii) NCCO (dec+ul) which replaces the suppression mechanism with a decrement mechanism that reduces the number of individuals as much as possible. The experiments on dynamic benchmark functions in the two and five dimensions based on the Moving Peaks Benchmark (MPB) demonstrate that NCCO (sup+ul) and NCCO (dec+ul) successfully track multiple optima in dynamic multimodal functions with lower errors than Multi-swarm Quantum Particle Swarm Optimization (mQSO) as the conventional method. Specifically, (1) NCCO (dec+ul), followed by NCCO (sup+ul) and NCCO, decreases Fitness Error and Distance Error, and (2) NCCO (dec+ul) also maintains the smallest number of individuals during the search.

Existing prediction-based dynamic multi-objective evolution-ary algorithms (DMOEAs) face significant challenges in tracking dynamic Pareto optimal set (POS) for large-scale dynamic multi-objective optimization problems, including insufficient high-quality training data and the lack of prior knowledge about optimal solutions in the new environment. To address these issues, we propose a large language model (LLM)-based DMOEA. The proposed method jointly models environmental time-varying parameters and dynamic POS within a unified semantic space via LoRA-based fine-tuning. The fine-tuned LLM leverages parameter variation sequences from historical to new environments, together with historical POS sequences, as contextual information to accurately pre-dict the change directions of POS in a new environment. In doing so, this facilitates the evolution progress in finding the POS in the new environment. Additionally, we design an LLM reasoning-guided evolutionary search that narrows the search region from the large-scale search space to a promising local region, generating an initial population with both convergence and diversity for the new environment. Comprehensive exper-imental results on widely used dynamic benchmark problems demonstrate that LLM-DMOEA significantly outperforms state-of-the-art algorithms, particularly in large-scale scenar-ios.

Most practical engineering applications are often formulated as interval constrained multiobjective optimization problems (ICMOPs). In such problems, at least one objective or constraint contains interval parameters. With increasing constraints, the feasible region becomes narrow and fragmented. This greatly raises the difficulty of finding feasible solutions. Therefore, this paper proposes an interval-constrained multiobjective evolutionary algorithm integrating interval constraint clustering and transfer learning (ICCTL-ICMOEA). First, an interval constraint overlap degree is introduced to quantify constraint similarity, the Louvain algorithm is employed to cluster constraints into subclasses. Each constraint class is then combined with interval objectives to construct multiple subproblems. Next, within a multi-population evolutionary framework, a population state-aware search strategy is designed to independently evolve the main population and subpopulations. Following that, an interval constraint relaxation-guided knowledge transfer strategy is developed to migrate selected individuals from subpopulations to the main population, accelerating feasible region exploration. Finally, the proposed algorithm was tested on constructed benchmark ICMOPs and compared with a state-of-the-art interval-constrained multiobjective evolutionary algorithm and a constrained multiobjective evolutionary algorithm. Experimental results demonstrate the strong competitiveness of the proposed algorithm.

With the acceleration of urbanization, the demand for travel in urban areas and the number of private vehicles have risen sharply. However, the structure of the urban transportation network remains largely unchanged, and the road capacity is limited, unable to meet the increasing transportation demands, resulting in traffic congestion and further causing a series of negative impacts, such as time waste, aggravated environmental pollution, and frequent traffic accidents, ultimately leading to a low social benefit. Traditional travel planning usually only provides a few shortest path navigation suggestions, and the feedback of road conditions is often delayed, which easily leads travelers to make wrong traffic decisions, causing the traffic congestion to worsen. Reasonable traffic flow distribution can balance the traffic flow in the network and effectively alleviate traffic congestion. Therefore, the problem of traffic flow distribution has received widespread attention and research. This paper proposes a dual-population information interaction-driven evolutionary reinforcement learning algorithm, Dual-Drive-ERL, for solving the dynamic traffic flow distribution problem. This is the first time that the evolutionary reinforcement learning algorithm has been applied to the dynamic traffic flow distribution problem, and an innovative mechanism of dual-population information interaction-driven is proposed to enhance the exploration ability. The algorithm has shown superior performance to other comparison algorithms in experiments conducted in 45 different scenarios.

Closed-loop decision-making systems (e.g., lending, screening, or recidivism risk assessment) often operate under fairness and service constraints while simultaneously inducing feedback effects: decisions change who appears in the future, yielding non-stationary data and potentially amplifying disparities. We study online learning of a one-dimensional threshold policy from bandit feedback under demographic parity (DP) and (optionally) service-rate constraints. The learner observes only a scalar score each round and selects a threshold; reward and constraint signals are revealed for the chosen threshold only. We propose Optimistic Feasible Search (OFS), a simple grid-based method that maintains confidence bounds for reward and constraint residuals for each candidate threshold. At each round, OFS selects the threshold that appears feasible under confidence bounds and, among those, maximizes optimistic reward; if no threshold appears feasible, OFS selects the threshold minimizing optimistic violation. This design directly targets feasible high-utility thresholds and is particularly effective for low-dimensional, interpretable policy classes where discretization is natural.

We evaluate OFS on (i) a synthetic closed-loop benchmark with stable contraction dynamics and (ii) two semi-synthetic closed-loop benchmarks grounded in German Credit and COMPAS, constructed by training a score model and feeding group-dependent acceptance decisions back into population composition. Across all environments, OFS achieves substantially higher reward with dramatically smaller cumulative constraint violation than unconstrained and primal–dual bandit baselines. Moreover, OFS is near-oracle: its steady-state reward gap to the best feasible fixed threshold is below 0.003 across benchmarks. All experiments are reproducible and organized with double-blind friendly relative outputs.

Spiking Neural Networks (SNNs) offer a promising avenue for energy-efficient edge intelligence, yet their performance is often bottlenecked by the rigid, fixed dynamics of traditional neuron models like LIF and CuBa-LIF. These models typically rely on pre-defined exponential decay kernels, which fail to capture the diverse multi-scale temporal dependencies present in varying neuromorphic modalities. To overcome this limitation, we propose the Kernel-Learnable LIF (KL-LIF) neuron, a generalized model capable of end-to-end learning arbitrary temporal response kernels (Integrated and Reset kernels). To train deep networks composed of KL-LIF neurons efficiently, we derive the Spatio-Temporal Tilted Backpropagation (STT-BP) algorithm. By exploiting the duality between forward spike propagation and backward error flow, STT-BP establishes a tilted gradient path via kernelized error gradients, enabling unified optimization of synaptic weights and temporal kernels without complex state unrolling. Extensive experiments on neuromorphic vision (DVS-Gesture, DVS-CIFAR10) and spiking audio (SHD, SSC) datasets demonstrate that STT-BP significantly outperforms state-of-the-art methods. Notably, our approach reveals distinct optimal temporal scales for different modalities—shorter kernels for vision and longer kernels for audio—achieving top-tier accuracy.

Muscle synergies are widely hypothesized as an organizational principle for controlling highly redundant musculoskeletal systems. However, most existing synergy analyses rely on post-hoc statistical decompositions of recorded muscle activity, making it difficult to directly relate synergy function to closed-loop control. In this work, we propose a reinforcement learning-based neuromuscular control framework that embeds muscle synergies as explicit, policy-structured components within the control architecture. A synergy decoder maps low-dimensional modulation signals to high-dimensional muscle excitations, enabling synergy structure to emerge as an intrinsic part of the learned controller. Based on this, we introduce a decoder-based synergy analysis method using neural reconstruction and synergy ablation. We evaluate the proposed approach on a high-dimensional musculoskeletal torso balance task. Synergy function is probed through both open-loop replay ablation and closed-loop online ablation. The results show that reconstruction-based importance and functional importance are related but not equivalent. It suggests a division of labor among synergy pathways, with different synergy channels contributing to feasibility, terminal stabilization, and precision shaping. Overall, this work provides a control-relevant framework for studying the functional roles of muscle synergies and bridges biomechanical motor control with reinforcement learning. The experiment video can be found in: https://sites.google.com/view/syn-decoder.

Inspired by the natural emergence of roles in biological systems for efficient coordination, role-based multi-agent reinforcement learning (MARL) methods have been proposed to enhance policy heterogeneity and cooperative capability in complex tasks. However, existing approaches typically rely on a fixed number of roles, whereas in practical multi-agent systems, the required role cardinality often varies across task stages and agent numbers, causing learned cooperative strategies to transfer poorly under varying agent numbers. In this paper, we propose SRD-DRC, a scalable role discovery framework with dynamic role cardinality. The core idea is to dynamically aggregate a global context from the local interaction histories of all agents via an attention mechanism, and to adaptively estimate the number of roles required at each cooperation stage, thereby eliminating reliance on a predefined role cardinality. Based on this estimation, SRD-DRC employs contrastive learning to acquire discriminative role representations that capture behavior patterns evolving across cooperation stages. Moreover, role representations are integrated at both the policy and value levels by parameterizing role-conditioned heterogeneous layers and guiding centralized value decomposition to enable structured credit assignment. Experimental results on the StarCraft Multi-Agent Challenge (SMAC) benchmark demonstrate that SRD-DRC achieves superior performance across diverse cooperative scenarios and exhibits effective zero-shot scalability.

Diffusion models achieve outstanding image-generation performance, yet practical deployment requires continual knowledge expansion under distribution shift. Most continual learning studies focus on memory constraints (limited access to old data). In practice, however, the primary bottleneck is often compute: GPU time and the number of allowable training iterations (optimizer steps) are limited. To improve the final generation quality with a limited number of iterations, we propose Diffusion-model One-iter Forgetting-based Replay (DOFR). Our idea comes from empirical observation that the loss increase after a single iteration on a new task (one-iteration forgetting) serves as a low-cost yet informative proxy for long-term forgetting. Based on this insight, DOFR estimates the forgetting risk at each diffusion timestep in a simple closed form, enabling replay budgets to be allocated to classes with higher forgetting risk. Experiments with Stable Diffusion v1.5 show that, under the same compute budget, DOFR improves overall generation quality over simple replay baselines, yielding a better compute-performance trade-off.

Explainable Artificial Intelligence is critical for trustworthy time series forecasting, yet recent studies reveal its vulnerability to adversarial manipulations. Existing attacks, primarily designed for images, fail to preserve temporal semantics and often require white-box access, rendering them ineffective against non-differentiable methods like LIME and SHAP. To bridge this gap, we propose the Iso-surface Curvature Maximization Attack, a geometric attack method grounded in the insight that explanation instability is intrinsically governed by the curvature of the prediction manifold. Instead of targeting specific explainer outputs, ICMA directly maximizes the local curvature of the decision boundary under a spectro-temporal constraint. This constraint strictly confines perturbations to the high-frequency residual subspace, preserving the underlying trend and seasonality. Extensive experiments on real-world datasets across diverse architectures confirm that maximizing this geometric objective significantly destabilizes explanations, empirically verifying that high curvature is indeed the primary driver of explanation fragility.

Hyperspectral image (HSI) classification faces a persistent challenge: modeling long-range semantic context while preserving fine-grained local structures. Existing paradigms often struggle to balance these needs; Convolutional Neural Networks (CNNs) are limited by local receptive fields, whereas Transformer-style attention can mix irrelevant regions and overlook the sequential nature of spectral signatures. Moreover, treating HSI as a unified volume often blurs the distinct roles of the spectral and spatial axes. To address these issues, this paper proposes the Hyperspectral Retention Network (HSRNet), which factorizes feature learning into two stages: a spectral retention encoder that captures continuous band correlations in a pixel-wise spectral token sequence, followed by a spatial retention encoder that aggregates geometric context over the flattened spatial tokens. A distance-decayed retention operator with head-wise learnable decay factors controls the extent of contextual mixing, allowing different heads to emphasize short-range interactions or preserve per-token details, while a lightweight local context enhancement module injects explicit 2D neighborhood cues. Experiments on Pavia University and Houston 2013 datasets demonstrate the effectiveness of HSRNet, achieving Overall Accuracies of 98.9% and 98.3%, respectively, and producing more coherent classification maps than recent state-of-the-art baselines.

Network pruning is critical for supporting the efficient deployment of convolutional neural networks (CNNs) on resource-limited TinyML hardware. While pruning has proven effective in many cases, semantic segmentation models remain challenging to prune for deployment to TinyML devices due to their large output feature maps and consequent memory accesses, necessitating more fine-grained compression strategies. We propose TinySegPrune: a multi-objective pruning framework that parameterizes the pruning rate of each block in the network. We employ a latency predictor that estimates latency from block-wise pruning rates for fast hardware performance evaluation. We validate the effectiveness of our framework by pruning two different segmentation networks for a TinyML processor. Our framework finds a pruned model that reduces latency by 30.3% with only 1% accuracy reduction, outperforming benchmark methods that achieve only 14.7% latency reduction---a 2.1X larger improvement. Leveraging latency feedback, our framework effectively identifies critical blocks, reducing their latency by 30% to 50%, and significantly improving overall efficiency.

Although large-scale pre-trained models have revolutionized audio classification, adapting them to downstream few-shot tasks remains challenging. Existing methods face a dilemma: conventional few-shot adaptation requires resource-intensive gradient updates, while zero-shot inference often suffers from semantic misalignment in text prompts. To address these issues in deployment-constrained scenarios, we propose Audio-Guided Adaptive Prompt Ensembling (AGAPE), a training-free framework for few-shot audio classification. Specifically, AGAPE first utilizes limited support shots to construct optimal class-specific text prototypes, thereby enhancing text-prompt-based classification. Subsequently, the final classification is achieved by fusing the similarity scores of the test sample against both these optimized text prototypes and the audio prototypes. Furthermore, we introduce the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) for Test-Time Adaptation (TTA), enabling efficient model adaptation without gradient updates. Experiments on ESC-50 and FSD50K demonstrate that AGAPE significantly outperforms the baseline. Ablation studies further validate the effectiveness of our methods.

Radar-based human pose estimation enables privacy-preserving motion tracking for ambient intelligence, yet the noisy nature of radar sensing makes uncertainty quantification essential. We present RadProPoser, an end-to-end probabilistic framework that predicts three-dimensional body joints with per-joint uncertainties from raw radar tensor data. Using a variational encoder-decoder with spectral attention that fuses real and imaginary radar components across temporal frames, we model aleatoric uncertainty through learnable Gaussian and Laplace distributions. Trained on a new benchmark dataset with optical motion-capture ground truth, our method achieves 6.425 cm mean per-joint position error. The model outputs per-joint aleatoric uncertainties, and isotonic recalibration yields calibrated total uncertainty with expected calibration error of 0.027. Since spectral attention operates on individual radar tensor components, extending to multi-radar configurations requires only concatenating additional input streams. On the HuPR benchmark with dual orthogonal radars, this achieves 5.042 cm MPJPE. The framework runs at approximately 89 frames per second (FPS) on an NVIDIA RTX 3090, exceeding the 15 Hz radar frame rate.

Brain connectome analysis has long been a focus in neuroscience and neuroimaging research. However, many recent approaches focus heavily on fMRI and ROI-based modeling approaches, often overlooking key structural features and connectivity patterns of white matter tracts. In this work, we present a novel heterogeneous graph framework that incorporates structural features of white matter tracts, representing both white matter clusters and anatomical regions as nodes with edges capturing their spatial intersection. This approach allows us to model both structural integrity and connectivity, along with regional context, for more biologically meaningful patterns. Using GraphSAGE for feature aggregation and multi-head Graph Attention Networks (GAT) for interpretability, our proposed architecture achieves an AUC of 0.891 for predicting gender on the Human Connectome Project 1200 dataset, matching the performance of other existing models and validating our graph construction while using significantly fewer training data. Furthermore, we extract attention scores, identifying key white matter tracts and anatomical regions involved in sex differentiation, grounding model predictions in neuroanatomical relevance. Our results highlight the importance of structural information and pave the way for more biologically informed models.

Since the introduction of Masked Autoencoders, various improvements to masking strategies have been explored. In this paper, we rethink masking strategies for audio representation learning using masked prediction-based self-supervised learning (SSL) on general audio spectrograms. While recent informed masking techniques have attracted attention, we observe that they incur substantial computational overhead. Motivated by this observation, we propose dispersion-weighted masking (DWM), a lightweight masking strategy that leverages the spectral sparsity inherent in the frequency structure of audio content. Our experiments show that inverse block masking, commonly used in recent SSL frameworks, improves audio event understanding performance while introducing a trade-off in generalization. The proposed DWM alleviates these limitations and computational complexity, leading to consistent performance improvements. This work provides practical guidance on masking strategy design for masked prediction-based audio representation learning.

Models with similar predictive accuracy are often assumed to learn comparable internal representations. We challenge this assumption in continuous-depth networks by showing that the numerical solver used to discretize Neural Ordinary Differential Equations (Neural ODEs) systematically alters the geometry of learned representations, even under matched predictive performance. Using a third-order Neural ODE framework, we study how Runge-Kutta (RK4) and extended-stability solvers (ESRK-15 and ESRK-21) shape latent trajectories in models trained within typical run-to-run accuracy variability ($\approx$2\%) on CIFAR-10 and CIFAR-100. Despite identical integration horizons and comparable validation accuracy, the resulting representations exhibit markedly different geometric structure, as quantified by curvature and jerk statistics derived directly from the latent dynamics. Our results show that solver stability properties and internal amplification act as implicit inductive biases that regulate representational smoothness and higher-order dynamics. These findings demonstrate that numerical integration is not merely an implementation detail, but a structural component of the model that influences how continuous-depth networks encode and regularize information.

Long-term time series forecasting(LSTF) is crucial for applications such as energy management and traffic forecasting, yet remains challenging due to complex multi-scale temporal dependencies, including long-term trends and mixed periodic patterns. Recently patch-based methods improve local temporal modeling by partitioning sequences into subseries, but fixed-length patching struggles to adapt to diverse and varying periodic structures in real-world data. To address this issue, we propose PP-Mixer, a fully MLP-based framework for LTSF. PP-Mixer employs a Fourier-guided multi-scale pooled patch embedding to adaptively construct multi-scale representations aligned with dominant temporal periods. A multi-scale decomposition mixer module then models and integrates seasonal and trend components across scales, while a dual-dimension mixing block enables efficient temporal modeling and feature interaction within each scale. Experiments on eight real-world benchmarks show that PP-Mixer consistently outperforms Transformer-based and MLP-based baselines, with ablation and visualization analyses confirming the effectiveness of its period-aware and multi-scale design.

Few-Shot Segmentation (FSS) aims to recognize and segment novel classes in images using only a limited number of annotated examples. Recently, increasing efforts have explored incorporating foundation models (e.g., DINOv2 and SAM) into FSS. However, existing pixel-wise matching approaches based on DINOv2 features often rely solely on foreground-constrained modeling, resulting in less discriminative similarity responses and suboptimal dense prompts. Moreover, relying solely on dense prompts remains inadequate for effectively guiding the Segment Anything Model (SAM) family in constraining object boundaries and low-confidence regions. To address these challenges, we propose EasyFSS, a SAM2-based few-shot segmentation framework that explicitly enhances dense prompts quality while introducing complementary sparse prompts. Specifically, based on multi-level DINOv2 features, we construct a Dual-Polarity Hierarchical Block (DPHB), extending foreground-constrained modeling to foreground–background polarity matching across semantic hierarchies. This design significantly enhances the discriminability of pixel-wise similarity responses and alleviates background-induced ambiguity. Furthermore, we design a Sparse Prompt Generation (SPG) module that leverages global prototypes and multi-scale spatial context to generate high-confidence sparse prompts, effectively constraining boundaries and low-confidence regions. Finally, dense and sparse prompts collaboratively guide the SAM2 decoding process, enabling accurate and reliable segmentation predictions. Extensive qualitative and quantitative experiments demonstrate that EasyFSS achieves state-of-the-art performance, with 1-shot mIoU improvements of 3.2% and 3.9% on PASCAL-5^i and COCO-20^i, respectively. Our code is publicly available at https://github.com/stoneoceam/EasyFSS.

This paper presents an energy-efficient event-driven framework for cardiac arrhythmia classification that combines Level-Crossing Analog-to-Spike Converter (LC-ASC) sensing with a lightweight convolutional neural network (CNN) mapped onto a resistive compute-in-memory (CIM) system-on-chip. The proposed approach transforms the sparse, asynchronous output of the LC-ASC into compact spatio-temporal spike-frame representations, enabling effective feature extraction using standard CNN architectures while significantly reducing data volume. Evaluated on the MIT-BIH Arrhythmia Database, the proposed event-driven CNN achieves high classification accuracy and strong robustness to class imbalance, outperforming Nyquist-sampled CNN baselines in Macro-F1 score while requiring substantially fewer computations. To further improve efficiency, the network is deployed on an RRAM-based CIM processor comprising 288 parallel crossbar arrays, enabling highly parallel vector–matrix multiplication with minimal data movement. Hardware evaluation demonstrates sub-millisecond end-to-end latency per heartbeat and microjoule-level energy consumption, achieving orders-of-magnitude improvements compared to both desktop and edge GPU platforms. These results validate the proposed framework as a practical, real-time, and energy-efficient solution for near-sensor intelligence in wearable ECG monitoring systems.

Rapidly training reliable machine learning models from limited data remains highly desirable in real-world time-series classification (TSC) applications. Reservoir Computing (RC), originating from dynamical system modeling and characterized by fixed high-dimensional nonlinear projections with lightweight training, has emerged as a promising paradigm for small-sample TSC. In RC-based TSC models, a significant challenge lies in how the classifier can effectively utilize the rich temporal representations generated by the reservoir to produce accurate classification results. In this paper, we propose a novel subsampling strategy for reservoir states, termed Quantile-based Positive Proportion (QPP). Unlike existing subsampling strategies, QPP transforms continuous reservoir representations into binary representations using optimizable quantile thresholds and constructs features by computing the proportion of positive activations. Experiments are conducted on 13 small-sample TSC benchmark datasets from the UCR archive using two representative RC models. Comparative results against existing subsampling strategies show that the proposed QPP consistently achieves superior average rankings across datasets, which highlights the high effectiveness and generalization capability of QPP with respect to RC models in dealing with small-sample TSC tasks.

Prototype-Based Networks (PBNs) provide intrinsic interpretability in image classification by learning characteristic traits of objects as evidence to support predictions. While most existing PBNs operate in Euclidean embedding spaces, recent work shows that representing images as linear subspaces provides robust and interpretable models. However, current subspace-based prototype methods are deterministic and do not explicitly account for (i) the manifold’s (in this case Grassmann) geometry during optimization; (ii) the uncertainty in the model's decisions.

We propose a probabilistic Learning Vector Quantization (LVQ) classifier, defined directly on the Grassmann manifold. Our approach models predictive uncertainty, which is important for human-centric applications such as in healthcare and justice, and employs geometry-aware optimization that respects the manifold’s intrinsic structure. We also use gradient-based explanations to show which regions of the image drive the classifier’s decisions. Experiments on three benchmark datasets demonstrate competitive performance compared to previous PBNs. These also show that the classifier aggregates evidence from different image regions to make its final prediction and identifies potential sources of uncertainties in the image.

To enable rigorous, interpretable analysis of hot-rolling surface-quality mechanisms, we propose an end-to-end framework for spatiotemporal registration and transferable causal discovery. It targets two central bottlenecks in industrial data-driven modeling: physically-inconsistent representations and regime-sensitive causal discovery under distribution shift. First, a physics-driven nonuniform spatiotemporal warping operator maps heterogeneous time series to a unified spatial domain, resolving spatiotemporal mismatches. Second, the Transferable Industrial Causal Engine (TiCE) formulates a progressive bilevel training paradigm that embeds mechanistic constraints and jointly activates an Independence-Constrained Domain-Invariant regularization with a Hierarchically Adaptive Fusion Weighting mechanism to resist structural drift under mixed regimes. Third, learned causal graphs guide Asymmetric SHAP attributions to preserve causal consistency in mechanistic explanations. Benchmarks and ablations on TE and UF-IMD demonstrate robustness to distributional shift and the necessity of modular synergy. In hot rolling, the framework reveals a binary antagonistic mechanism for surface defects: temperature-dominated oxidation kinetics versus multi-stage descaling suppression, thereby providing credible decision support for root-cause diagnosis in complex processes.

Bridge structure monitor relies on precise crack detection, yet current deep learning surveillance methods rely heavily on a large number of crack images as dataset, which are difficult to collect in complex and dangerous underwater environments. To alleviate this problem, we propose a training-free diffusion framework that enables efficient and high-quality underwater crack generation. Specifically, we first apply the denoising diffusion implicit models inversion to extract crack and underwater representations. We propose a structure preservation module that leverages global and local information from crack representations to enable the preservation of fine-grained crack details, mitigating structural loss. In parallel, we propose a style integration module to integrate underwater domain information and prevent content leakage issue. To better align the underwater domain style, we design a color adaptation module that dynamically performs adaptive instance normalization, which adjusts the color distribution of the translated result and further optimizes the visual realism of underwater crack images. Extensive experiments show that the proposed framework achieves a significant improvement on generating high fidelity underwater cracks. Code is available at https://github.com/LiHaoyang0517/High-Fidelity-Underwater-Crack-Generation

Progressive structured pruning is a widely adopted strategy for compressing convolutional neural networks, where model capacity is gradually reduced through multiple prune–fine-tune stages. While most existing methods focus on defining effective importance criteria for convolutional filters, the order in which they are removed is typically fixed and the effect of pruning order is largely unexplored. In this work, we investigate different pruning order strategies for structured pruning of convolutional filters. Specifically, we introduce two alternative scheduling policies, tail-based and center-based, that modify the order of filter removal while preserving the same final pruning budget. Unlike classic approaches, which remove the least important filters first, the proposed scheduling strategies modify the pruning order of the components. In this work, we demonstrate the significant impact of pruning scheduling through extensive experiments with ResNet and ResNeXt architectures on CIFAR-100 and ImageNet-1K datasets. In particular, center-based scheduling consistently outperforms the standard progressive strategy, with larger gains at higher pruning rates and on more challenging datasets. Additional analyses show that these improvements are robust with a varying number of pruning stages and importance criteria. These results highlight how changing pruning scheduling is a simple yet effective method to improve structured pruning performance.

Long-term navigation for service robots faces critical challenges like the accumulation of odometry drift and sensor error, which progressively degrade 2D maps and renders traditional path planning algorithms (e.g., A*, RRT*, DiPPer, ViT-A*) ineffective over time. To address this, we propose a user-friendly, interactive framework that eliminates the reliance on globally consistent maps. Our approach integrates visual perception with Large Language Models (LLM) to interpret user commands via text or voice. Instead of relying on a drift-prone global map, the system generates a sequential action plan based on local visual cues and egocentric geometric instructions. These action plans are executed sequentially, allowing the robot to navigate known and unknown environments safely. By resetting localization relative to immediate targets, our framework effectively works with a minimum accumulation drift strategy, ensuring accurate, efficient, and collision-free navigation without the maintenance overhead of traditional mapping. Experiments on real-world and simulated data have shown significant improvements over other methods. Our code and dataset will be made public.

Automated sequence labeling of regulatory documents is crucial for automated compliance checking (ACC) and domain knowledge graph (KG) creation, yet is hindered by a ``high-cost, low-resource'' dilemma. Public, high-quality labeled datasets for building codes are scarce and difficult to reuse across regions, making large-scale annotation expensive. Although large language models (LLMs) show strong few-shot capabilities, multi-round, clause-by-clause annotation with high-performance APIs remains prohibitively costly, whereas affordable models are less reliable. To resolve the cost-performance dilemma, we propose LMABC, a low-cost, low-resource LLM Multi-Agent framework for Automated Sequence Labeling in the Building Codes. Rather than relying on a single powerful model, LMABC coordinates multiple diverse, affordable LLM agents to achieve high-quality annotation through collective intelligence. The framework couples a task allocation mechanism with a two-phase labeling workflow consisting of parallel labeling and chain-based transfer labeling. Final results are obtained by a hybrid consensus ensemble that combines majority voting with dynamic weighting to support downstream KG creation. Experiments on building-code and general-domain benchmarks demonstrate that LMABC approaches high-performance models at a fraction of the cost, and achieves higher accuracy with fewer tokens than state-of-the-art (SOTA) methods on general tasks.

Zeroth-Order (ZO) optimization enables On-Device Learning (ODL) on NPU-equipped microcontrollers by estimating gradients through forward passes alone, bypassing the need for backpropagation primitives and reducing memory requirements. The number of gradient samples q critically affects training: insufficient samples produce noisy gradients that plateau early, while excessive samples consume more computational resources. However, finding an optimal q typically requires costly hyperparameter searches. This work introduces QScheduler, an adaptive algorithm that adjusts q based on training progress, and provides the first proof-of-concept of INT8 quantized on-device training on the STM32N6's Neural-ART NPU. Experiments on EuroSAT and STL-10 show that QScheduler matches well-tuned fixed-q configurations for both ResNet18 and MobileNetV2, without requiring prior q hyperparameter optimization.

Large Language Models (LLMs) offer powerful capabilities in code generation, natural language understanding, and domain-specific reasoning. Their application to quantum software development remains limited, in part because of the lack of high-quality datasets both for LLM training and as dependable knowledge sources. To bridge this gap, we introduce PennyLang, an off-the-shelf, high-quality dataset of 3,347 PennyLane-specific quantum code samples with contextual descriptions, curated from textbooks, official documentation, and open-source repositories. Our contributions are threefold: (1) the creation and open-source release of PennyLang, a purpose-built dataset for quantum programming with PennyLane; (2) a framework for automated quantum code dataset construction that systematizes curation, annotation, and formatting to maximize downstream LLM usability; and (3) a baseline evaluation of the dataset across multiple open-source and commercial models, including ablation studies, all conducted within a retrieval-augmented generation (RAG) pipeline. Using PennyLang with RAG substantially improves performance: for example, Qwen 7B’s success rate rises from 8.7% without retrieval to 41.7% with full-context augmentation, and LLaMa 4 improves from 78.8% to 84.8%, while also reducing hallucinations and enhancing quantum code correctness. Moving beyond Qiskit-focused studies, we bring LLM-based tools and reproducible methods to PennyLane for advancing AI-assisted quantum development.

Portfolio management allocates wealth across assets to maximize returns. Reinforcement learning (RL) is promising, but it must balance risk-controlled decisions with rapid adaptation to changing markets. Offline RL trains reliably on fixed historical data using supervised-style objectives, yet its performance is limited by dataset quality and coverage. Online RL can adapt through exploration, but it requires costly real-time interaction with the environment. We introduce O2O-TP, an offline-to-online Transformer RL framework that pretrains on historical data and then refines online with low computational overhead. O2O-TP embeds a Dirichlet policy head to produce simplex-constrained portfolio weights, enabling deterministic or stochastic execution. Across four stock markets, O2O-TP achieves cumulative wealth up to 3.3 times higher than established finance and RL baselines, with low variance between random seeds. Online refinement keeps efficient, accounting for only 14% to 23% of total runtime. These results indicate that the proposed framework enables effective and computationally efficient adaptation in dynamic financial markets.

Sequential decision making paradigms have recently emerged in Multi-Agent Reinforcement Learning (MARL) to address the coordination limitations of simultaneous execution. However, existing approaches typically employ fixed or observation-driven heuristic orderings, overlooking the intrinsic dependency of decision priority on agent roles. In this paper, we propose the Role-Guided Prioritized Multi-Agent Transformer (RPMAT), a novel framework that optimizes action generation order by leveraging explicit role semantics. Specifically, we design a Role-Guided Cross-Attention (RGCA) mechanism that queries global observation features using learned latent role representations to dynamically quantify the decision urgency of each agent. This mechanism allows high-contribution roles to preemptively broadcast their intentions, thereby guiding the team's sequential execution. Theoretically, we reinterpret this order optimization as anterior structural credit assignment, where agents are ranked based on their predicted potential contribution to the global value function. Extensive experiments on SMAC, Google Research Football, and Multi-Agent MuJoCo benchmarks demonstrate that RPMAT achieves consistent superiority over state-of-the-art baselines in both heterogeneous and homogeneous scenarios, validating the effectiveness of role-driven coordination.

Recent advances in Earth observation satellites have enabled high-resolution imagery for Internet of Remote Things (IoRT) applications, but such imagery often contains privacy-sensitive information that may be exposed during transmission or processing. Addressing privacy protection directly on orbit is challenging due to the strict computational, energy, and communication constraints of Low Earth Orbit (LEO) satellite systems and their limited ground-station contact windows. In this paper, we propose a Distributed On-orbit Privacy Protection framework DOPP that integrates lightweight neural detection and inpainting with inter-satellite collaborative computing. Each satellite employs a quantized YOLO-based detection model to identify privacy-sensitive regions and generate compact binary masks, which enable non-overlapping image slicing and parallel inpainting across neighboring satellites via inter-satellite links. A fine-tuned Mobile Inpainting Generative Adversarial Network (Mi-GAN) reconstructs masked regions with low computational overhead while preserving visual consistency. To ensure timely downlink transmission, a visibility-aware aggregation strategy dynamically selects the satellite with the most imminent ground-station access to assemble and transmit the processed imagery. Simulations on real-world remote sensing datasets demonstrate that DOPP achieves effective privacy preservation with high accuracy, while significantly reducing processing latency and energy consumption , resulting in a 15–20% lower Energy-Delay Product (EDP) compared with existing onboard cooperative baselines. These results highlight the feasibility of lightweight neural privacy protection for real-time, resource-constrained satellite imaging systems.

Event-based neuromorphic sensors generate sparse and asynchronous streams whose reliability as neural inputs depends not only on global event statistics, but critically on the preservation of underlying spatio–temporal structure. In practice, noise corruption, localized signal loss, or aggressive preprocessing may disrupt structural continuity while leaving global statistics largely unchanged, yielding inputs that appear statistically plau- sible yet are structurally invalid. Such mismatches can lead to abrupt performance collapse or silent failure in neural systems. We propose a structure-preserving local–global graph processing framework for reliability diagnosis in event-based neural systems. Rather than serving as a denoising or enhancement module, the proposed processing stage is explicitly formulated as a di- agnostic intervention operator that applies controlled, structure- aware graph transformations to probe neural sensitivity under progressive structural degradation. The method jointly captures local neighborhood coherence and global geometric organization, enabling targeted structural interventions without directly opti- mizing task accuracy. To support systematic analysis, we introduce a structure- centric evaluation protocol based on geometry-driven objects, multi-scale motion settings, and standardized degradation mech- anisms, explicitly decoupling structural corruption from global event statistics. Experimental results demonstrate that structural integrity degrades significantly earlier than global statistical indi- cators and serves as a reliable predictor of abrupt failure in both spiking and non-spiking models. Compared with purely local or purely global strategies, the proposed local–global formulation exhibits superior robustness under severe structural degradation, revealing failure modes that remain invisible to conventional preprocessing and task-driven evaluation.

Punctuation restoration improves ASR (AutomaticSpeech Recognition) readability, however streaming ASR requires online decisions with limited future context. In streaming ASR, the system predicts punctuation incrementally, which makes generation-based approaches prone to latency and alignment failures under boundary-wise evaluation. This paper proposes a non-autoregressive scoring method (no free-form generation) that preserves the input transcript and makes a decision at each word boundary. Our method compares punctuation insertion hypotheses against a no-insertion baseline under a bounded K-subword-token lookahead, and calibrates decisions using a weight α and a validation-calibrated threshold τ (no parameter updates during inference). On IWSLT 2017, our scoring method achieves a 4-class macro F1 of 0.893 in the no-fine-tuning setting (validation-calibrated, K=2) and 0.937 after fine-tuning (K=2), outperforming the prompt-based baseline (0.566) and a fine-tuned ELECTRA baseline (0.913) under the same lookahead budget. We analyze the impact of the lookahead budget through ablation studies on K.

Accurate short-term wave forecasting is essential for ocean monitoring and for supporting stakeholder decision-making. This paper proposes a neural framework for one-hour-ahead prediction of significant wave height, mean wave period, and mean wave direction using the ERA5 dataset. The proposed approach uses a local 3 × 3 spatial grid to capture the essential neighbourhood dynamics of the wave and wind fields. The information from this grid, sampled over the two previous time steps, forms a 108-dimensional input vector that represents local spatio-temporal interactions. An autoencoder is then used to compress this representation into a 32-dimensional latent space, which is subsequently mapped to the target variables using a multilayer perceptron. The performance of the model is evaluated in a spatially disjoint test region. The results show high predictive accuracy, with coefficients of determination above 0.9959 for all target variables, and low absolute errors. The proposed framework provides a computationally efficient alternative for short-term wave forecasting suitable for integration into a digital twin of the ocean.

Indoor mobile robots require a comprehensive understanding of their surroundings to navigate and interact safely in cluttered environments. This typically involves multiple complementary perception tasks, such as scene classification, semantic segmentation, and panoptic segmentation, under real-time constraints. While recent mask-based segmentation methods already shift instance separation into the model, many existing approaches still remain computationally expensive and have mainly been studied in RGB-only settings. In this paper, we propose IRSAFormer, a Transformer-based approach for indoor scene analysis that follows the efficient Encoder-only Mask Transformer (EoMT) paradigm by injecting query tokens into a largely unmodified Vision Transformer (ViT) backbone and using lightweight task heads directly on top of it. To address the limitations of RGB-only perception, we analyze RGB-D incorporation strategies. Our results indicate that depth provides clear gains when pre-training is limited, whereas strong large-scale pre-training narrows the gap between RGB and RGB-D. Beyond closed-set prediction, we investigate distilling text-aligned embeddings as a more flexible output representation for language-driven downstream applications. We find that scene-level embeddings distill reliably, whereas query token-level embeddings remain more challenging. Overall, IRSAFormer achieves state-of-the-art performance on NYUv2, SUN RGB-D, and ScanNet V2 while maintaining high efficiency, reaching 26.8 FPS on an NVIDIA Jetson AGX Orin with a ViT-Base backbone.

Real-world deployment of multi-agent reinforcement learning (MARL) systems is fundamentally constrained by limited compute, memory, and inference time. While expert policies achieve high performance, they rely on costly decision cycles and large-scale models that are impractical for edge devices or embedded platforms. Knowledge distillation (KD) offers a promising path toward resource-aware execution, but existing KD methods in MARL focus narrowly on action imitation, often neglecting coordination structure and assuming uniform agent capabilities. We propose resource-aware Knowledge Distillation for Multi-Agent Reinforcement Learning (KD-MARL), a two-stage framework that transfers coordinated behavior from a centralized expert to lightweight, decentralized student agents. The student policies are trained without critic, relying instead on distilled advantage signals and structured policy supervision to preserve coordination under heterogeneous and limited observations. Our approach transfers both action-level behavior and structural coordination patterns from expert policies with supporting heterogeneous student architectures, allowing each agent's model capacity to match its observation complexity, which is crucial for efficient execution under partial or limited observability along with limited onboard resources. Extensive experiments on SMAC and MPE benchmarks demonstrate that KD-MARL achieves high performance retention while substantially reducing computational cost, retaining over 90% of expert performance while reducing computational cost by up to 28.6× FLOPs. The proposed approach preserves expert-level coordination through structured distillation, enabling practical MARL deployment across resource-constrained onboard platforms.

Chromatic aberration (CA) remains a common optical artifact in digital imaging, causing spatial misalignments between color channels and resulting in visible color fringing. In this paper, we propose TriadNet, a novel hybrid convolutional-transformer architecture that formulates CA detection as a multimodal feature learning problem for accurate analysis of real-world photographs. TriadNet integrates three complementary modality-specific pathways: an edge-centric pathway that captures structural, high-frequency information where CA manifests most prominently, a channel-aware attention mechanism that models chromatic inter-channel discrepancies, and an edge-modulated feature processing strategy that facilitates effective multimodal feature fusion. Experimental evaluation on a custom dataset of real-world images demonstrates that TriadNet outperforms several baselines, achieving $F_1$-scores exceeding 0.98 and robust generalization across varying aberration strengths. The proposed method offers a lightweight yet effective solution for automatic CA detection within modern image processing pipelines.

This paper introduces Text-Conditioned Flow Matching Policy (TC-FMP), a continuous-time generative imitation learning framework that leverages natural language as a lightweight yet effective guidance signal for visuomotor control. TC-FMP learns an instruction-conditioned velocity field that guarantees the smooth evolution of trajectories from noise toward the demonstrated action distribution. To mitigate mode collapse and improve conditional fidelity, classifier-free guidance is incorporated by randomly dropping textual conditioning during training. At inference time, guided and unguided vector fields are combined to produce a scaled velocity field, which is integrated using a second-order Heun solver to generate action sequences. The resulting action trajectories are executed in a receding-horizon control scheme, enabling closed-loop execution over long horizons. Language is fused with visual and proprioceptive inputs via FiLM modulation, introducing negligible architectural overhead while allowing natural language prompts to steer task-relevant behaviors. TC-FMP is first evaluated on the planar PushT benchmark, and then to demonstrate its robustness and practical utility; on a long-horizon, contact-rich bimanual peg-in-hole insertion task under varying noise conditions. Experimental results show that increasingly specific language instructions consistently improve both average and best-case returns compared to diffusion-based and flow-matching baselines. These findings highlight that even single-task visuomotor control can substantially benefit from language-driven guidance, improving robustness, controllability, and performance.

Chiari type I malformation (CM-I) is characterized by the caudal herniation of cerebellar tonsils through the foramen magnum. Diagnosis traditionally relies on measuring tonsillar position relative to the McRae line, a manual process prone to inter-observer variability. While AI has advanced skeletal landmark prediction in CT, applying these methods to MRI remains challenging due to low osseous contrast, creating a dependency on radiation-based imaging. In this work, we present CMInet, the first fully MRI-based framework integrating cerebellar segmentation with direct osseous landmark regression. This approach eliminates the need for supplementary CT scans, streamlining clinical workflows. Leveraging a rare, hardware-limited paired dataset from specialized upright MRI, our model effectively captures subtle gravity-dependent structural changes. Our framework delivers an objective tool for anatomical assessment, reducing manual workload while establishing a critical normative baseline to distinguish physiological mobility from pathological herniation.

Federated learning (FL) faces key challenges in communication efficiency and privacy protection. Most existing studies primarily address one of these aspects, while comprehensive security mechanisms often conflict with efficiency requirements. Binary neural networks (BNNs) can substantially reduce computation and communication costs in FL, but aggressive binarization typically leads to severe performance degradation. In this work, we propose BiSec-FL, a secure and efficient FL framework that integrates binary aggregation with homomorphic encryption (HE). Specifically, we introduce Scale and Residual Compensated Binary Aggregation (SR-BinAgg) to mitigate quantization loss in binary aggregation, and adopt a selective secure aggregation scheme based on multi-key HE to protect full-precision parameters. Experiments show that BiSec-FL achieves stable convergence across multiple datasets and settings, reducing single-round communication costs by over 96% on average compared with FedAvg, while incurring significantly lower cryptographic overhead than fully encrypted FL systems.

Chaotic convective flows arise in many real-world systems, such as microfluidic devices and chemical reactors. Stabilizing these flows is highly desirable but remains challenging, particularly in chaotic regimes where conventional control methods often fail. In this work, we improve the practical feasibility of controlling chaotic flows by Reinforcement Learning (RL), focusing on Rayleigh-Bénard Convection (RBC), a canonical model for convective heat transport. In particular, our goal is to robustly reduce the convective heat transfer. Although RL has shown promise for control in laminar flow settings, its ability to generalize and remain robust under chaotic and turbulent dynamics is not well explored, despite being critical for real-world deployment. Hence, to enhance generalization and sample efficiency, we introduce domain-informed RL agents that are trained using Proximal Policy Optimization across diverse initial conditions and flow regimes. We incorporate domain knowledge in the reward function via a term that encourages Bénard cell merging, as an example of a desirable macroscopic property. In laminar flow regimes, the domain-informed RL agents reduce convective heat transport by up to 33%. In chaotic flow regimes, the agents still achieve a significant 10% reduction, which is markedly better than the conventional controllers used in practice. We compare the domain-informed to uninformed agents: Our results show that the domain-informed reward design results in steady flows, faster convergence during training, and generalization across flow regimes without retraining. Our work demonstrates that elegant domain-informed priors can greatly enhance the robustness of RL-based control of chaotic flows, bringing real-world deployment closer.

Large language models (LLMs) have shown strong code generation ability, yet most benchmarks concentrate on high-resource languages such as Python, Java, and C++, leaving a gap in evaluating LLM performance on low-resource and legacy languages. Languages including COBOL, Fortran, Ada, and Prolog remain in financial infrastructure, aerospace systems, scientific computing, and symbolic reasoning, but are supported by far less training data and lack widely accepted benchmarks. To fill this gap, we present CodeFORGE, an automated benchmark synthesis framework tailored to data-scarce programming languages. Our method uses LLMs to generate programming tasks together with corresponding test suites through a designed pipeline. CodeFORGE adopts a model-cascading verification strategy: multiple LLMs attempt each generated task in sequence, leveraging iterative feedback from earlier failures to enable cross-validation and ensure benchmark reliability. Only tasks that satisfy automated checks inside isolated Docker environments are included in the final benchmark. With this framework, we build CodeFORGE-Bench, a comprehensive benchmark spanning 12 low-resource languages across diverse paradigms, including procedural, functional, and logic programming. Extensive experiments evaluating state-of-the-art LLMs reveal performance gaps relative to high-resource languages, offering insights into current limitations when LLMs face linguistically diverse and data-scarce code generation settings.

Applying deep learning to instance-aware re- identification of slate tiles and extraction site classification can improve production efficiency and quality control in the slate tile industry. These tasks are particularly important for handling natural materials where visual variability can make manual inspection costly and error-prone. We present a lightweight, hybrid deep learning approach that combines image matching and classification within a single framework. The system integrates a feature-matching branch based on XFeat with a MobileNetV3- based classification branch. The XFeat branch, combined with a LightGlue matching head, improves instance matching performance by +15.4% AUC. For classification, features from both backbones are shared and fused, resulting in a +10.9% accuracy improvement over a standard MobileNetV3 model. Our approach is evaluated on a newly created industrial dataset consisting of 2,610 slate tile images from six extraction sites. The results demonstrate the effectiveness of the proposed approach for object re-identification and classification in an industrial setting.

Semi-structured $N\!:\!M$ sparsity enables hardware acceleration on modern GPUs, but its use in heterogeneous multi-agent reinforcement learning (MARL) remains largely unexplored. Existing parameter-sharing methods either ignore hardware constraints or encourage policy homogenization, while strict $N\!:\!M$ sparsity introduces additional challenges due to non-stationary activations, discrete constraints, and agent heterogeneity. We propose PATH-NM (Pattern-Aware Transformation for Hardware-Efficient $N\!:\!M$ Sparse MARL), a sparsity-aware framework that learns agent-specific $N\!:\!M$ sparse subnetworks on top of a shared backbone. PATH-NM reformulates pruning as pattern distribution learning and combines dynamic activation-aware scoring, differentiable pattern transformation, and orthogonal diversity regularization. This design supports end-to-end optimization, adapts to non-stationary RL dynamics, and explicitly promotes structural diversity across agents. Experiments on MPE and SMACv2 show that PATH-NM consistently outperforms representative parameter-sharing baselines in both sample efficiency and final performance, while remaining fully compatible with hardware-accelerated $N\!:\!M$ sparse inference.

Cross-Project Defect Prediction (CPDP) involves building models using data from external projects, called training projects, to predict modules from the target project. However, traditional CPDP methods suffer from the distribution shift between training and target projects that affects the model's performance. This paper proposes a novel CPDP framework that addresses this issue by proposing a two-stage multiple classifier system (MCS) selection scheme: one working at the project level and another at the module level. In the first stage, the framework evaluates multiple possible MCS configurations to find one that covers and generalizes well across multiple training projects. Consequently, the proposal is likely to obtain a diverse set of classifiers, each specialized in tackling software modules with distinct characteristics. The second selection stage operates at test time, selecting the most competent classifiers to predict each new module in the target project. Unlike previous approaches that apply the same classifiers to the entire target project, the proposed framework performs module-level model selection. This way, the system is more robust to changes in distributions between training and target projects because the selected set of classifiers is module-dependent. Our experimental results using 82 projects from four different CPDP benchmark datasets demonstrate that the proposed approach outperforms the state-of-the-art CPDP methods in most scenarios. The code, dataset, and further details about the proposed method are publicly available at https://github.com/jsaj/Multi_DES.

We present an FPGA implementation of Growing Neural Gas (GNG) on a Sipeed Tang Nano~9K (Gowin GW1NR-9C) as a bounded-memory hardware substrate for online self-organizing clustering. The design uses an FSM datapath with BRAM-based storage, integer fixed-point arithmetic (Q8/Q16), and a compact \emph{edge-cell graph encoding} that stores the upper-triangular adjacency as 8-bit age+1 values, avoiding dense adjacency matrices and dynamic memory allocation. The complete design is implemented without any soft-core processor or floating-point unit, and is evaluated on Two Moons and Concentric Circles benchmarks. The platform establishes a hardware foundation for deploying continual self-organizing clustering on resource-constrained edge devices.

Probabilistic forecasting for multivariate stepwise time series is difficult because observations are dominated by long stable regimes, while predictive changes occur sparsely and asynchronously across variables. We propose a dual-path neural controlled differential equation framework that disentangles persistent trend structure from event-driven updates in continuous time. A trend pathway summarizes regime-level dynamics via smoothing, whereas an event pathway sparsifies the primary channel by retaining endpoints and statistically significant change windows to form an irregular event sequence. The two representations are fused by a learned time-varying gate and mapped to predictive distribution parameters. We study Gaussian, Laplace, and Student-t likelihoods and evaluate forecast quality using the continuous ranked probability score together with calibration and sharpness metrics. Experiments show consistent improvements over classical and neural baselines, and indicate that heavy-tailed likelihoods improve distributional accuracy under regime shifts, while calibrated interval reliability remains subject to a sharpness trade-off.

Federated learning (FL) enables collaborative model training without centralizing raw data, but its performance can degrade under non-identically distributed (non-IID) client data and unreliable client updates. This paper introduces Φ-FL, a server-side aggregation strategy that augments Federated Averaging with trust fusion based on multiple reliability signals computed on a small server proxy set, robust magnitude control via interquartile-range (IQR) norm clipping, and interaction-sensitive reweighting using a 2-additive Choquet-inspired mechanism to promote complementary trusted updates. We evaluate Φ-FL on a suite of benchmark tabular classification datasets under Dirichlet-partitioned non-IID settings, using repeated paired runs. Compared to FedAvg, Φ-FL improves the mean test accuracy from 69.56% to 74.74%, achieving higher mean accuracy on 9 out of 12 datasets. In relation to more advanced baseline model, FLTrust, the proposed method Φ-FL achieves average accuracy higher by 1.7 percentage point, while the median accuracy increases by 3.7 percentage point. The significance of the obtained results is confirmed by statistical tests.

Conventional federated learning assumes that greater learner participation improves training performance, by leveraging abundant, independently generated local data. However, in federated reinforcement learning (FRL) for unmanned aerial vehicle (UAV) teams in hazardous environments where experience generation is severely constrained by safety considerations, energy limitations, and mission duration, this assumption may break. This work introduces Experience-Constrained Hierarchical Federated Reinforcement Learning (EC-HFRL), a framework in which clusters act as federated learning agents, while multiple intra-cluster learners represent parallel learning resources that reuse a shared experience pool. We show that increasing participation does not necessarily improve learning performance. Instead, learning performance is strongly associated with experience reuse strategy and the dominance of key analytically identified gradient transition experiences within a cluster. In particular, minibatch size primarily determines effective replay exposure, while higher intra-cluster participation increases reuse level. Empirical results demonstrate that the commonly observed performance regimes are strongly associated with the structure of the learning signal, rather than federated aggregation effects, clarifying the limited and secondary role of learner participation in experience-constrained FRL.

Backpropagation underpins the success of artificial neural networks, yet it has critical limitations such as backward locking and global error propagation. The Forward-Forward algorithm has been proposed as a more biologically plausible method that replaces the backward pass with an additional forward pass. Nevertheless, the Forward-Forward algorithm significantly trails backpropagation in accuracy, and its optimal form requires multiple forward passes at inference time, resulting in low inference efficiency. In this work, the Forward-Forward algorithm is reformulated within a similarity-learning paradigm. This proposed algorithm is named Forward-forward Algorithm Unified with Similarity-based Training (FAUST). FAUST incorporates similarity-based objectives and, by caching estimates of class centroids, achieves single-pass inference. On matched MLP and CNN architectures, FAUST yields improved accuracy over Forward-Forward baselines and narrows the gap to backpropagation across MNIST, Fashion-MNIST, and CIFAR-10. Applying FAUST to ResNet-18 yields competitive results on ImageNette.

Accurate decoding of motor imagery electroencephalography (MI-EEG) is fundamental to brain–computer interface applications, yet remains challenging due to low signal-to-noise ratio and amplitude fluctuations. Existing approaches often underutilize spatial relationships and struggle to model long-range temporal dependencies. We propose AwareNet, an end-to-end spatio-temporal EEG decoding framework that employ a Bidirectional Temporal Convolution layer to capture both past and future contextual information, enriching temporal feature representations. Next, we design a lightweight Coordinate-aware Spatial Gating module that leverages prior knowledge of actual electrode positions to model spatial relationships. Furthermore, we introduce the Channel-aware Transformer module to capture global dependencies. CAT is an eeg-friendly attention mechanism which replaces conventional dot-product attention with cosine similarity and incorporates per-head learnable temperature scaling, enabling more stable and adaptive temporal modeling under noisy, small-sample conditions. Extensive experiments demonstrate that AwareNet consistently outperforms state-of-the-art baselines on multiple datasets. Visualization analyses further confirm that AwareNet yields more discriminative and neurophysiologically consistent representations, with our CAT module surpassing conventional MSA in training stability, decoding performance, and interpretability. Overall, AwareNet provides accurate, stable, and interpretable MI-EEG decoding, demonstrating strong potential for practical BCI deployment.

Abstract—Event cameras provide sparse and asynchronous event streams with high temporal resolution, enabling human action recognition under fast motion and challenging illumina- tion conditions. However, effective modeling of event streams for event-based human action recognition remains challenging due to polarity imbalance, timestamp jitter, and the inherent mismatch between structural reasoning and transient temporal dynamics. To address these challenges, we propose CGSSN, a coupled graph–spiking dual-path framework that jointly models structure-aware spatial relations and spike-driven temporal dy- namics. CGSSN introduces a unified polarity-aware Poisson tem- poral encoding (UPPTE) module to enhance spike stability and robustness against temporal jitter, and adopts parallel graph and spiking pathways to capture complementary structural and tran- sient cues. In addition, a lightweight cross-branch alignment and gated fusion strategy is incorporated to improve representation consistency across heterogeneous pathways. Experimental results on five public event-based human action recognition benchmarks demonstrate that CGSSN consistently outperforms existing CNN- based, Transformer-based, and spiking-based methods, achieving state-of-the-art performance on multiple datasets. Comprehensive Ablation studies further verify the effectiveness of each proposed components.

Neural network potentials (NNPs) accelerate molecular dynamics (MD), but stable NNP-MD requires training data with broad configurational coverage, including rare high-energy states. Conventional knowledge distillation pipelines fine-tune a foundation teacher on density functional theory (DFT) and then use the fine-tuned teacher to generate training data via NNP-MD simulations for training a lightweight student. We show this order can harm MD: fine-tuning steepens the teacher potential energy surface, traps trajectories in low-energy basins, and collapses high-energy coverage. We therefore reverse the workflow. We keep the teacher non-fine-tuned as a generalist model to sample diverse NNP-MD configurations labeled with teacher energies and forces, distill a lightweight student on these soft targets, and then fine-tune only the student on a small DFT-labeled subset. On polyethylene glycol (PEG) and Li10GeP2S12 (LGPS) materials using MatterSim as the generalist teacher, the student achieves near-experimental property accuracy while reducing expensive DFT labeling by 10× (PEG) and 5× (LGPS) relative to representative active learning baselines. It also yields large inference gains (e.g., 82× speedup for a 3,100-atom PEG system) and avoids teacher memory bottlenecks. In PEG, We further show that a fine-tuned-teacher pipeline can match static validation error yet fail in MD due to missing high-energy coverage, showing that configurational coverage is as important as static accuracy for NNP‑MD.

The reconstruction of indoor scenes is a critical task for embodied robotics, where accurate perception of geometry is essential for navigation, scene understanding, and spatial reasoning. Although recent 3D Gaussian Splatting (3DGS) methods show promising results in novel view synthesis, their performance degrades in textureless regions and complex indoor layouts, resulting in incomplete and noisy reconstructions. In this paper, we introduce 2DGS-Room, a novel method leveraging 2D Gaussian Splatting for high-fidelity indoor scene reconstruction. Specifically, we introduce a seed-guided strategy to control the distribution of 2D Gaussians through dynamic growth and pruning, while depth and normal priors enhance geometry in both detailed and low-texture regions. Furthermore, multi-view consistency constraints improve structural robustness across views. Extensive experiments demonstrate that our method achieves state-of-the-art performance in indoor scene reconstruction.

Medical Visual Question Answering (Med VQA) presents a unique challenge demanding rigorous precision and zero hallucinations. Unlike general VQA tasks, interpreting medical imagery requires rigorous multi-scale feature understanding and multi-step diagnostic reasoning. Existing approaches, predominantly relying on single-agent Multimodal Large Language Models, often suffer from “tunnel vision”, focusing on salient features while missing subtle yet critical anomalies and lack robust self-verification mechanisms. Furthermore, the static inference strategies employed by current models lack the flexibility to adapt to varying question complexities, leading to inefficient resource scheduling. Inspired by how a prism decomposes light to reveal spectral details, we propose PRISM-Med (Perception, Routing, Inspection, Self-correction Multi-agent Collaboration System for Medical VQA), a modular framework. Our system features three key innovations: (1) a Two-Stage Diagnostic Architecture mimicking clinical “Triage-to-Consultation” with Organizer-based dynamic routing; (2) a Clinically Grounded Multi-Agent Framework with four specialized experts to mitigate tunnel vision; (3) a Dynamic Inference Optimization Strategy based on Proximal Policy Optimization (PPO) for adaptive generation configuration. PRISM-Med achieves state-of-the-art performance among open-source models on PMC-VQA, MedXQA, MMMU and OmniMedVQA benchmarks, outperforming baselines by approximately 3.68% on average. The code is available at https://gitee.com/zhou-jie021125/med.git.

Automatic speech recognition (ASR) systems often degrade on accented speech because acoustic-phonetic and prosodic shifts induce a mismatch to training data, making labeled accent adaptation costly. Text-centric pseudo-label filters (e.g., perplexity) can favor fluent hypotheses that remain acoustically misaligned; using them as supervision during fine-tuning amplifies errors. We present a reference-free selection pipeline for accent adaptation under a transductive, label-free protocol: unlabeled target speech drives pool reduction and percentile thresholds without oracle transcripts. We optionally apply target-aware FLMI preselection, then decode several hypotheses per utterance and score them with speech–text alignment in a frozen SONAR embedding space and predicted word error rate (WER), treating multimodal consistency as a proxy for pseudo-label quality under accent-induced mismatch. A simple percentile rule selects the training subset. In-domain, selecting ∼1.5k utterances from a 30k pool achieves 10.91% WER, close to 10.45% with 30k supervised labels. With a mismatched cross-domain candidate pool, consistency-filtered subsets avoid degradation from unfiltered pseudo-labels under strong accent shift; matched-hour experiments on a stronger ASR backbone further confirm gains over random sampling and recent selection baselines.

Collaborative perception faces critical pragmatic challenges under bandwidth constraints, particularly in guaranteeing perception performance while overcoming spatial-temporal imperfections such as sensor misalignment and asynchrony. Although recent fully sparse approaches have improved computational efficiency, they remain constrained by local receptive fields that limit long-range reasoning, lacking intelligent communication redundancy reduction, and still exhibit limited robustness to real-world noise. To address these challenges, we propose CoSTr, a Collaborative Sparse Transformer framework that operates natively on sparse feature representations through integrated sparse convolution and transformer. Our approach introduces a robust spatial-temporal attention mechanism that explicitly compensates for pose errors and communication delays during feature fusion. To minimize communication overhead, we propose a mutual information-based criterion that operates directly on sparse features to select and transmit only the most critical information. Extensive experiments on OPV2V and V2XSet demonstrate that CoSTr not only achieves state-of-the-art perception performance and communication efficiency, but also significantly improves the robustness against spatial-temporal disturbances, establishing a new benchmark for pragmatic collaborative perception.

Single-image scanning electron microscope (SEM) depth estimation plays a critical role in semiconductor three-dimensional metrology. However, existing end-to-end approaches often struggle to maintain stable generalization performance in complex structural scenarios composed of multiple geometric primitives. To address this issue, we propose a Divide-and-Conquer style structured modeling approach for single-image SEM depth estimation. The proposed method explicitly decomposes complex scenes into multiple local structural regions and shares depth prediction parameters across regions, thereby guiding the model to learn reusable local structure--depth mapping relationships and reducing its reliance on global layout statistics. The proposed approach is systematically validated on a simulation-generated SEM dataset. The training data include both basic structural primitives and their compositions forming SRAM-like structures, while the testing phase further evaluates the model’s prediction stability under variations in structural compositions. Experimental results demonstrate that, compared with conventional end-to-end models, the proposed method achieves more stable and reliable depth estimation performance in complex compositional scenarios, validating the effectiveness of structured modeling strategies in improving the generalization capability of single-image SEM depth estimation.

Explainable Recommender System (ExRec) provides transparency to the recommendation process, increasing users' trust and boosting the operation of online services. With the rise of large language models (LLMs), whose extensive world knowledge and nuanced language understanding enable the generation of human-like, contextually grounded explanations, LLM-powered ExRec has gained great momentum. However, existing LLM-based ExRec models suffer from profile deviation and high retrieval overhead, hindering their deployment. To address these issues, we propose Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation (REXHA). Specifically, we design a hierarchical aggregation based profiling module that comprehensively considers user and item review information, hierarchically summarizing and constructing holistic profiles. Furthermore, we introduce an efficient retrieval module using two types of pseudo-document queries to retrieve relevant reviews to enhance the generation of recommendation explanations, effectively reducing retrieval latency and improving the recall of relevant reviews. Extensive experiments demonstrate that our method outperforms existing approaches by up to 12.6% w.r.t. the explanation quality while achieving high retrieval efficiency.

Dementia subtyping from routine scalp EEG could enable low-cost decision support, but deep models often ignore 10–20 electrode geometry and can overfit small clinical cohorts. We introduce graph-harmonic scalpograms, a geometry-aligned representation that projects 19-channel EEG onto low-frequency scalp graph harmonics and converts harmonic activity into compact time-frequency images. Using Open-Neuro ds004504 (N = 88, eyes-closed; AD/FTD/CN), recordings are segmented into 8 s windows (75% overlap), projected onto the first K = 12 non-DC graph modes, transformed over 1–45 Hz using F = 60 log-spaced bins, and encoded as Band-RGB scalpograms (δ+θ, α, β+γ) under strict subject-wise 80/10/10 splits to prevent leakage. Models are trained with class-weighted cross-entropy, geometry-preserving spectro-temporal augmentations, and post-hoc temperature scaling to improve probability calibration; evaluation is performed at the subject level by averaging window posteriors with 95% BCa bootstrap confidence intervals. A lightweight CNN trained from scratch achieves the best subject-level performance (balanced accuracy 0.83 ± 0.07,macro-F1 0.847 ± 0.06, AUROC 0.95 ± 0.05), outperforming transfer baselines such as ResNet-34 (0.65 ± 0.05 /0.667 ± 0.04 / 0.90 ± 0.03). Overall, aligning the model inductive bias with 10–20 sensor geometry supports accurate, well-calibrated subject-level dementia subtyping using lightweight CNNs that are promising for small clinical EEG cohorts.

Mixture-of-Experts (MoE) models provide a structured approach to combining specialized neural networks and offer greater interpretability than conventional ensembles. While MoEs have been successfully applied to image classification and semantic segmentation, their use in object detection remains limited due to challenges in merging dense and structured predictions. In this work, we investigate model-level mixtures of object detectors and analyze their suitability for improving performance and interpretability in object detection. We propose an MoE architecture that combines YOLO-based detectors trained on semantically disjoint data subsets, with a learned gating network that dynamically weights expert contributions. We study different strategies for fusing detection outputs and for training the gating mechanism, including balancing losses to prevent expert collapse. Experiments on the BDD100K dataset demonstrate that the proposed MoE consistently outperforms standard ensemble approaches and provides insights into expert specialization across domains, highlighting model-level MoEs as a viable alternative to traditional ensembling for object detection. Our code is available at \url{https://github.com/KASTEL-MobilityLab/mixtures-of-experts/}.

Green wave guidance is an important application in intelligent transportation. This paper proposes a green wave guidance strategy based on cross-network collaboration. First, a cross-network collaboration framework is introduced to optimize multi-intersection green wave guidance, incorporating an improved Ising model for road pass-through rate assessment. Then, an improved ant colony algorithm based on road pass-through rate distribution is developed to generate vehicle trajectory guidance and speed recommendations, with a dynamic trajectory update strategy to enhance information usability. Experimental results show that this method effectively evaluates traffic conditions and improves vehicle passage efficiency.

Accurate and systematic evaluation of mobile agents can significantly advance their development and real-world applicability. However, existing benchmarks for mobile agents lack practicality and scalability due to the extensive manual effort required to define task reward signals and implement corresponding evaluation codes. To this end, we propose AutoEval, an autonomous agent evaluation framework that tests a mobile agent without any manual effort. First, we design a Structured Substate Representation to describe the UI state changes while agent execution, such that task reward signals can be automatically generated. Second, we utilize a Judge System that can autonomously evaluate agents’ performance given the automatically generated task reward signals. By providing only a task description, our framework evaluates agents with fine-grained performance feedback to that task without any extra manual effort. We implement a prototype of our framework and validate the automatically generated task reward signals, finding over 93% coverage to human-annotated reward signals. Moreover, to prove the effectiveness of our autonomous Judge System, we manually verify its judge results and demonstrate that it achieves 94% accuracy. Finally, we evaluate the state-of-the-art mobile agents using our framework, providing detailed insights into their performance characteristics and limitations.

The Strong Lottery Ticket Hypothesis (SLTH) posits that high-performing subnetworks can be extracted from randomly initialized neural networks using binary pruning masks, or supermasks. This paper introduces the Trichromatic Strong Lottery Ticket Hypothesis (T-SLTH), a unified generalization of the SLTH that decomposes supermasks into three conceptually orthogonal components: connectivity, sign, and magnitude. This trichromatic decomposition unifies existing SLTH variants, enables new configurations, and clarifies the connection between supermask-based training and quantization-aware training. The proposed framework is evaluated on image classification benchmarks with ResNet architectures, achieving state-of-the-art accuracy-compression tradeoffs among supermask-based models. ResNet-50 is compressed 38x on CIFAR-100 and 25x on ImageNet without training any weights. Additional experiments across datasets, architectures, and modalities demonstrate that trichromatic supermasks act as robust structural priors. These results position partially random networks trained via supermasks as a compelling alternative to conventional weight optimization for efficient inference systems.

Referring Image Segmentation (RIS) aims to segment objects based on natural language expressions. Existing methods leverage Multimodal Large Language Models (MLLMs) to interpret text descriptions and localize objects, while mask generation follows two paradigms. The first approach represents segmentation masks using polygon coordinates, which inherently suffers from precision errors. The alternative approach integrates MLLMs with Segmentation Foundation Models (SFMs). Combining two models with differing output mechanisms can lead to optimization inefficiencies and hinder the effective propagation of information. In this work, we propose DMR-Seg, an image segmentation method that integrates mask generation within MLLM. We use the modified VQVAE to discretize binary segmentation masks into sequences of tokens. By training the MLLM to predict these sequences autoregressively, the model is endowed with the capability to achieve image segmentation. To reduce decoding dependency caused by autoregressive generation, we further designed a training strategy that proactively injects token noise. Experimental results on multiple datasets demonstrate that our model outperforms existing segmentation methods. Furthermore, ablation studies confirm the effectiveness of the proposed training strategy.

We propose GP-Adapter, a training-free framework that augments CLIP (Contrastive Language-Image Pre-training) with Gaussian Process (GP) uncertainty modeling for few-shot classification and out-of-distribution (OOD) detection. While CLIP achieves strong zero-shot recognition, it yields deterministic similarity scores and offers limited uncertainty information, which is critical under distribution shift and data scarcity. GP-Adapter constructs modality-specific, class-wise one-class GPs on top of frozen CLIP embeddings using an RBF kernel for image features and a linear kernel for text prompts and fuses their predictive statistics to produce a variance-aware confidence score for OOD detection. The method requires no fine-tuning of the CLIP backbone and relies only on a small K-shot cache and lightweight hyperparameter selection, with memory cost scaling as O(CK^2) for C classes and K shots. Experiments on ImageNet and multiple OOD benchmarks show that GP-Adapter provides competitive few-shot performance and consistently improves OOD detection when combined with prompt-learning baselines, highlighting the complementarity between GP-based uncertainty modeling and prompt learning. Overall, our results suggest that integrating probabilistic inference with large pre-trained vision-language models can improve reliability in low-data and distribution-shifted settings. Code is available at https://github.com/tms-byte/GP-Adapter

This paper introduces Q-SARIMA, a hybrid quantum-classical architecture for 10-day meteorological forecasting in precision agriculture. By integrating SARIMA structural robustness with Variational Quantum Circuits (VQC), the model maps qubit expectation values to statistical coefficients using the gradient-free COBYLA optimizer. Evaluation on real-world Brazilian datasets via walk-forward validation shows that while classical SARIMA provides superior stability, Q-SARIMA is competitive for high-volatility series like wind speed. These results highlight NISQ-era challenges, such as convergence instability and cost surface roughness. This study establishes a methodological baseline for hybrid quantum-classical integration, offering critical insights into the trade-offs between quantum expressivity and statistical reliability in agricultural monitoring systems.

Hyperparameter Optimization (HPO) is essential for building high-performing ML/DL models, yet conventional optimizers often struggle in high-dimensional spaces where evaluations are costly and progress is diluted across many low-impact variables. We propose Greedy Importance First (GIF), an importance-aware scheduling strategy that uses a small-sample warm start to estimate hyperparameter importance, forms importance-based groups, allocates trials proportionally, and retains a full-space fallback. We evaluate GIF under fixed evaluation budgets on five anisotropic analytic functions ($d\!\in\!\{5,10,30,50\}$), Bayesmark, and NAS-Bench-301 (33D). On the higher-dimensional benchmarks, GIF reaches better incumbents with faster convergence than TPE, BOHB, Random Search, and Sequential Grouping. On Bayesmark, where the effective dimensionality is smaller, GIF remains competitive but the margins are smaller. Ablation studies show that importance estimation, proportional allocation, and the fallback step all contribute to the gains. We also verify that the HIA component recovers the intended anisotropy on the analytic benchmarks. These results suggest that GIF is a simple and plug-compatible way to improve sample efficiency in high-dimensional HPO.

The Visual Geometry Grounded Transformer (VGGT) marks a significant leap forward in 3D scene reconstruction, as it is the first model that directly infers all key 3D attributes (camera poses, depths, and dense geometry) jointly in one pass. However, this joint inference mechanism requires global attention layers with extremely long sequences that causes a significant latency bottleneck. In this paper, we propose blockwise clustered attention (BC attention) to accelerate the global attention layers in VGGT. By limiting the clustering within HW-friendly neighborhood blocks, BC attention reduces the computation overhead of query clustering as well as the costly data movement between on- and off-chip memory. This enables BC attention to scale to long sequences and deliver practical latency improvements on GPUs. Moreover, we introduce a hashing hyperplane calibration method and a threshold-based error compensation method to reduce clustering errors efficiently, which is a bottleneck in the current clustered attention mechanism. Overall, our experiments on GPU demonstrate that calibrated BC attention accelerates the global attention layers by 2.10-2.63× and the whole backbone by 1.77-2.35× with negligible loss (1%) for large scenes. With a small performance loss (< 5%), calibrated BC attention further achieves a 2.26-2.87× latency improvement on the global attention layers and a 1.90-2.55× improvement on the backbone.

Human Activity Recognition (HAR) in smart environments enables assisted living, health monitoring, and home automation. Traditional HAR methods often rely on handcrafted features or shallow models that do not generalize across different sensor layouts. Sequence-based deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), improve temporal modeling but typically treat sensor streams as independent channels. As a result, these models do not explicitly encode spatial relationships between sensors or the structural constraints imposed by the environment, limiting their ability to model sensor interactions driven by the environment's physical structure. We study Graph Convolutional Networks (GNNs) for HAR using ambient sensor data, modeling sensors as nodes and their spatio-temporal interactions as edges. Our graph construction uses domain-informed room adjacency and temporal co-activation, with shared node and edge encoders. We evaluate a Graph Convolutional Network (GCN), a Graph Attention Network (GAT), and a hybrid GAT–Long Short-Term Memory model (GAT-LSTM) on two CASAS smart homes (HH111 and HH101). Sparse, domain-informed graphs consistently outperform fully connected graphs, and longer temporal windows improve robustness. GAT-LSTM attains the best classification performance, while GCN provides the most efficient baseline. Overall, combining spatial attention with temporal modeling yields a strong framework for HAR while highlighting remaining challenges for multi-resident activities and data-driven room adjacency.

Vertical loop compatibility—predicting the perceptual compatibility of two music loops layered together in a mix—has seen limited progress due to the absence of a large-scale, high-fidelity, labeled dataset. We introduce the Advanced Recombinant Transformer (ART). To train ART, we curate a large dataset of aligned positive pairs by rendering studio-quality stems from multi-instrument MIDI. We construct matched-condition negatives using score-based mining with calibrated thresholds based on a blinded musician audit. For representation, we replace log-mel features with frozen embeddings from the Music Understanding model with large-scale self-supervised training (MERT), which capture timbral nuances and provide cues for higher-level musical structure. An order-invariant head scores pairwise compatibility. Experimental results show a 10.1% improvement in compatibility classification over the previous state-of-the-art. Extensive subjective tests (N = 107) on production-grade, human-composed audio loops find seamlessness and creativity ratings comparable to human-curated pairings. Samples are available at: https://loopcompatibility.github.io/demo

Medical image segmentation is fundamental for clinical diagnosis, yet developing deep learning models remains challenging due to the scarcity of pixel-level annotations. Few-shot segmentation offers a promising solution by leveraging only a few labeled examples, but existing methods struggle with the support-query misalignment problem—a single support image often fails to represent the diverse appearances of target organs across patients. In this paper, we propose GenSAM, a novel training-free framework that addresses this challenge by leveraging diffusion models for test-time adaptation. Our key insight is that query-guided style injection can synthesize support views that are visually closer to the query domain while maintaining anatomical consistency, effectively bridging the support-query gap without additional annotations. GenSAM comprises three components: (1) Generative Support Expansion (GSE), which synthesizes diverse augmented views using query-guided diffusion-based style transfer; (2) Multi-center Prompting, which generates positive and negative point prompts from similarity maps to guide SAM2; and (3) Consistency-Aware Ensemble, which aggregates predictions weighted by confidence scores to filter unreliable results. Experiments on three medical imaging benchmarks demonstrate that GenSAM significantly outperforms existing training-free methods and achieves competitive performance with training-based approaches, highlighting the potential of combining generative and discriminative foundation models for medical image analysis.

Human Pose Estimation (HPE) provides a structured representation of human motion and has become a key component in a wide range of applications, spanning from medical monitoring and rehabilitation to sports biomechanics and consumer-oriented applications such as dance and yoga instruction. While 2D pose estimation is sufficient for some of these use cases, others require 3D pose information to be effective. Despite significant recent progress in HPE, achieving real-time 3D estimation with low latency is still a challenging problem. This challenge may be addressed through the use of event cameras — novel, bio-inspired, visual sensors which react to changes in the intensity values in the scene, asynchronously, at pixel level. This sensing modality is extremely efficient, as it removes redundancy at the sensor level, and is well suited for the design of high-speed, robust artificial vision algorithms, including HPE. In this paper, we combine MoveEnet, a HPE system based on event cameras which performs real-time high-frequency 2D pose estimation, with PoseFormerV2, a transformer-based 2D-to-3D pose lifting model to develop an end-to-end system for monocular real-time 3D HPE. Our method not only performs more accurately and faster than existing HPE models for event cameras, but is also competitive with RGB models while being substantially faster. Our final system has an end-to-end latency of 30ms, a real-time frequency of 38Hz and a 3D MPJPE of 85.7mm on event-Human 3.6M dataset. Code for the end-to-end real-time pipeline can be found at https://github.com/event-driven-robotics/HPE-event-poseformer.

Multimodal heterogeneous graphs are increasingly used to model complex systems combining diverse entity types, relations, and data modalities. While multimodal graph learning has shown strong empirical performance, existing methods often rely on uniform cross-modal alignment or indiscriminate fusion, which can induce negative transfer when modalities are imbalanced, sparse, or semantically incompatible. In this paper, we propose a self-supervised, transferability-sensitive framework for multimodal heterogeneous graph neural networks (S$^2$T-MMHGNN). Instead of enforcing symmetric alignment, our approach leverages directional predictive self-supervision to explicitly estimate inter-modality transferability and uses this signal to regulate multimodal fusion. This design selectively amplifies complementary modalities while suppressing unreliable cross-modal interactions. We provide theoretical analysis establishing collapse avoidance and monotone fusion behavior under increasing modality incompatibility. Extensive experiments on five real-world benchmarks demonstrate consistent improvements over state-of-the-art heterogeneous, multimodal, and self-supervised baselines, particularly under noisy and sparse multimodal settings.

In Knowledge Tracing (KT) tasks, to address the black-box characteristic of knowledge mastery, recent work has proposed some interpretable methods to represent knowledge mastery. Educational theories suggest that learners apply knowledge concepts across different cognitive levels. Therefore, cognitive levels should be represented in the unit of knowledge mastery, which is overlooked by existing studies. We propose the concept of "cognitive ability" as the unit of knowledge mastery to represent cognitive levels, and propose a Cognitive Ability-based Knowledge Tracing (CogKT) method. In CogKT, the relationships among cognitive abilities are modeled using cognitive ability graphs and Graph Neural Network (GNN), while the temporal dynamics of cognitive abilities is captured using LSTM network enhanced by cognitive level embedding. Evaluated on a real-world dataset comprising 2393 students, CogKT outperforms baseline models on performance metrics AUC and ACC. In addition, cognitive abilities exhibit a significant influence on test outcomes, indicating that they contribute to enhancing the interpretability of knowledge mastery representation.

Neural ordinary differential equations (ODEs) provide a principled continuous-depth formulation, but their practical use is often limited by slow and numerically fragile training due to the weaknesses of standard ODE solvers. We introduce extended-stability Runge--Kutta methods, explicit fixed-step solvers that admit much larger step sizes without numerical blow-up in negative-real-axis dominated regimes than classical schemes such as RK4, or adaptive methods such as Dormand--Prince.

By enlarging the real-axis stability region, these solvers traverse a fixed integration horizon with far fewer steps, yielding deterministic computation, reduced function-evaluation cost, and smoother optimization dynamics. Across CIFAR-10, CIFAR-100, and Tiny-ImageNet, extended-stability solvers match final accuracy (typically within 1--2\%) while improving gradient-flow conditioning, reducing gradient clipping rates (e.g., from 60--100\% to 0--25\% on CIFAR-10, depending on the horizon). The 15-stage variant achieves a $14\times$ wall-clock speedup over Dormand--Prince. Overall, our results show that stability-region geometry, rather than solver order or adaptivity, governs gradient-flow conditioning under fixed integration horizons, making extended-stability methods an effective tool for accelerating training and improving optimization conditioning in Neural ODEs

Few-shot medical image segmentation (FSMIS) has achieved notable progress, yet most existing methods mainly rely on semantic correspondences from scarce annotations while under-utilizing a key property of medical imagery: anatomical targets exhibit repeatable high-frequency morphology (e.g., boundary geometry and spatial layout) across patients and acquisitions. We propose RAP, a training-free framework that Retrieves, Adapts, and Prompt-Fits Segment Anything Model 2 (SAM2) for FSMIS. First, RAP retrieves morphologically compatible supports from an archive using DINOv3 features to reduce brittleness to single-support choice. Second, it adapts the retrieved support mask to the query by fitting boundary-aware structural cues, yielding an anatomy-consistent pre-mask under domain shifts. Third, RAP converts the pre-mask into prompts by sampling positive points via Voronoi partitioning and negative points via sector-based sampling, and feeds them into SAM2 for final refinement—without any fine-tuning. Extensive experiments on multiple medical segmentation benchmarks show that RAP consistently surpasses prior FSMIS baselines and achieves stateof-the-art performance. Overall, RAP demonstrates that explicit structural fitting combined with retrieval-augmented prompting offers a simple and effective route to robust training-free few-shot medical segmentation.

The growing availability of electronic health records (EHR) presents increasing opportunities for data-driven applications in healthcare. These records, however, are represented using disparate coding systems and often suffer from fragmentation and structural inconsistencies. Such data fragmentation can lead to substantial performance gaps across different predictive tasks. In this work, a framework is proposed in order to mitigate these artefacts and enable multi-task predictions. To address performance gaps, we introduce a concurrent loss that guides the learning process over fragmented entities. To address structural misalignment, we propose a neural message-passing algorithm that enables the integration of ontological information. Experiments conducted on cardiovascular disease (CVD) patients demonstrate that the proposed system overcomes prediction gaps across multiple clinical tasks, including laboratory tests, diagnoses, medications, and mortality prediction, performed concurrently. The results highlight the potential of ontology-aligned representation learning for high-precision decision-making in healthcare systems.

Hallucination in large language models (LLMs) contributes to the spread of misinformation and diminished public trust, particularly in high-risk domains. Among hallucination types, factuality is crucial, as it concerns a model's alignment with established world knowledge. Adversarial factuality, defined as the deliberate insertion of misinformation into prompts, tests a model's ability to detect and resist confidently framed falsehoods. However, existing work lacks high-quality, domain-specific resources for assessing model robustness under such adversarial conditions, and no prior research has examined the impact of injected misinformation on long-form text factuality.

To address this gap, we introduce \textit{AdversaRiskQA}, the first verified and reliable benchmark systematically evaluating adversarial factuality in Health, Finance, and Law domains. The benchmark includes two difficulty levels to test LLMs' defensive capabilities across varying knowledge depths. We propose two automated methods for evaluating the adversarial attack success and long-form factuality. We evaluate six open- and closed-source LLMs from the Qwen, GPT-OSS, and GPT families, measuring misinformation detection rates. Long-form factuality is then assessed under baseline conditions and with injected misinformation. Results show that after excluding meaningless responses, Qwen3 (80B) achieves the highest average accuracy at $\mathbf{94.7\%}$ across domains, while GPT-5 maintains consistently high accuracy. Performance scales non-linearly with model size, varies across domains, and gaps between difficulty levels narrow as models grow. Long-form evaluation reveals no significant correlation between the presence of misinformation and the model’s factual output. \textit{AdversaRiskQA} provides a valuable benchmark for pinpointing LLM weaknesses and advancing the development of more reliable models for high-stakes applications.

Depth estimation plays a pivotal role in robotics applications, where self-supervised approaches have emerged as promising solutions by leveraging abundant unlabelled real-world data. While existing methods achieve notable performance under static scene assumptions, their effectiveness significantly degrades in dynamic environments due to moving objects. To address this issue, we present ReloDepth, a novel method for self-supervised depth estimation in dynamic scenes. It tackles the challenge of dynamic objects from two key perspectives. First, within the self-supervised learning framework, we introduce a reprojection loss mask to explicitly localize regions likely contaminated by dynamic objects, thereby mitigating their corrupting influence on the reprojection-based supervisory signal at the loss level. Second, for multi-frame depth estimation, we propose a photometric gated cost volume strategy to mask redundant regions prior to cost volume, improving depth estimation robustness by steering the attention of the network toward photometrically informative regions. Furthermore, we propose a spectral entropy uncertainty module that incorporates spectral entropy to guide uncertainty estimation during depth fusion, effectively addressing issues arising from cost volume computation in dynamic environments. Extensive experiments on KITTI and Cityscapes datasets demonstrate that the proposed method consistently outperforms existing self-supervised depth estimation baselines.

Distributed Learning on structured data is a key and widely adopted approach in privacy-sensitive sectors, ranging from financial fraud prevention to federated healthcare diagnostics. While Gradient Boosted Decision Trees are widely used in centralized settings due to their efficiency, they are incompatible with standard gradient-based federated learning protocols because of their non-differentiability. Conversely, standard deep classifiers that are amenable to federated optimization often suffer from catastrophic forgetting under label distribution skew. Together, these limitations constitute the Federated Tabular Gap. In this work, to address this gap we propose Geodesic-SVDD, a framework for Distributed Spherical Prototype Learning that aggregates probabilistic manifolds rather than model weights. Specifically, for each client, we employ a Residual Spherical Encoder and Geodesic Latent Mixup that enforces strict hyperspherical constraints (∥ϕ(x)∥ ≡ 1), preventing the manifold collapse observed in linear approximations. Furthermore, leveraging this insight, we propose a novel One-Shot “Product of Experts” protocol that enables privacy-preserving aggregation without iterative synchronization, effectively handling disjoint class distributions. Finally, we mechanistically analyze the trade-off between interpolation strategies, demonstrating that while linear mixup offers an enticing computational shortcut, it destabilizes optimization gradients by orders of magnitude (109×). More broadly, our work showcases how differentiable manifold constraints can be leveraged to outperform state-of-the-art Federated methods by ≈ 20% on non-IID tabular benchmarks.

Diffusion models have emerged as the dominant framework for image generation, offering high fidelity samples and flexible conditioning. However, they may pose risks with regard to generating copyright-sensitive or inappropriate content that may be exploited by malicious actors. Machine unlearning emerged as a counteraction to this phenomenon, seeking to efficiently remove the influence of specific training data without sacrificing the general performance. Prior work on diffusion unlearning mostly targets text-to-image or class-conditional models, while image-to-image settings such as inpainting remain largely unexplored. In this work, we introduce DiF3CON (DiFFusion Forgetting via CONtinual unlearning), a continual unlearning framework for diffusion-based inpainting models. Inspired by regularization and stability mechanisms from multi-task learning, DiF3CON simulates iterative, efficient class-wise data removal, without compromising the performance of the base diffusion model. Extensive experiments on the Places365 dataset show that DiF3CON achieves unlearning stability over significantly larger forgetting data volumes, advancing this relatively underexplored area of machine unlearning for diffusion-based image-to-image models. Our code is available at https://github.com/Cata400/dif3con.

Aggregation operators form the fundamental set of tools used in information aggregation and fusion. The Joint Weighted Average (JWA) was recently proposed as a new aggregation operator that blends two different weighted aggregation approaches; source-based and evidence-based. Specifically, the JWA enables the systematic, joint consideration of the well-known linear weighted average (LWA) and the ordered weighted average (OWA) operators--and their associated weighting strategies. Here, we present a PyTorch-based toolkit to facilitate adoption of JWA aggregation in practice as well as to learn JWA weights from data. The latter is critical, not only for machine learning applications, but in particular for leveraging the explanatory power of the underlying weights, for example, in social science research. We integrate the operator learning approach with the established supervised learning framework that PyTorch provides, offering access to powerful optimization. We showcase and demonstrate the functionality of the toolkit. The open-source PyTorch-based toolkit opens the door for further development and accessible experimentation with the JWA framework and offers the only freely available approach for supervised learning in this setting. The toolkit is lightweight, extensible, and easy to use, allowing researchers and practitioners from other disciplines to leverage and advance JWA, associated techniques and research.

Recent years have seen growing interest in interval-valued (IV) survey response formats as an alternative to conventional single-valued (SV) scales, as they allow respondents to express uncertainty and imprecision as ranges. While IV data analysis methods have advanced, software support for applied IV survey analysis remains limited. To address this gap, this paper presents ISurvey, a free and open-source R toolkit for IV survey data analysis, designed for researchers and practitioners with diverse backgrounds. ISurvey supports datasets containing IV, SV, and single-choice items. It provides descriptive statistics under the Fréchet framework, visual descriptives, and survey psychometric assessment methods such as IV correlation for convergent validity and IV Cronbach's α for reliability. The toolkit also implements IV extensions of diagnostic methods used for actionable prioritization and benchmarking in survey research, including gap analysis, importance-performance analysis, and the customer satisfaction index. For illustration, we present a complete data analysis workflow using a UK rail customer survey, in which IV and SV responses were collected via a between-subject design. This provides a step-by-step tutorial that guides users from raw exports (from DECSYS, an open-source IV survey platform) to analysis, ready for reporting and decision-making.

In the field of quantum computing, Variational Quantum Algorithms (VQAs) combine quantum circuits with classical optimization to perform several tasks including classification and regression. Although gradient-based methods are commonly used for their training, these methods suffer from barren plateaus, noise sensitivity, and high computational costs. In this scenario, Evolutionary Algorithms (EAs) provide a gradient-free alternative for training VQAs enabling the discovery of high-quality solutions while requiring limited hardware resources. Unfortunately, the number of software tools aimed to the application of EAs for training of VQAs is limited and, moreover, they rarely exploits GPU acceleration. In order to bridge this gap, this work presents f-EVOVAQ, a Python package that extends the EVOVAQ framework by implementing vectorized evolutionary operators on GPU using CuPy arrays. The package enables efficient population-level optimization while remaining user-friendly for non-experts in evolutionary computation. Experiments on quantum classification tasks demonstrate significant reductions in optimization time and improved scalability, with limitations arising from GPU–CPU data transfers.

Nowadays, Large Language Models (LLMs) are widely included in many daily applications and services, where they play a crucial role in supporting users for automated decision-making tasks. As their adoption grows, so does their vulnerability to prompt injection attacks, which use natural language instructions to bypass safety constraints and also manipulate model behavior. To address these issues, the paper proposes a hybrid framework for prompt injection detection that combines domain-adaptive pretraining (DAPT) of RoBERTa encoder with a set of rule-based text patterns to detect suspicious adversarial interaction modes within prompt inputs. The proposed approach is evaluated through an empirical analysis across multiple prompt injection datasets. Experimental results show that the hybrid model outperforms text-only baselines and maintains reasonable performance under distributional shift. An ablation study is conducted to evaluate the individual component contribution.

Large Language Models (LLMs) can generate polished text, but they are not guaranteed to remain faithful to the underlying numbers. Fuzzy Linguistic Summarization (FLS) converts data into graded protoform statements with explicit fuzzy semantics and quality measures. We present FuzzyLinguistics, an open-source Python package for generating, simplifying, and comparing fuzzy linguistic summaries of tabular data. It offers a configuration-driven, Pydantic-validated workflow, vectorized protoform enumeration and scoring, graph-based redundancy reduction with optional compression, and differential summaries for comparing two datasets or model outputs under a shared linguistic vocabulary. Experiments on three minimal examples demonstrate deterministic outputs, strong redundancy reduction, and dataset comparisons, positioning FuzzyLinguistics as a practical grounding layer for explainable analysis and LLM-based reporting.

Support Vector Machines (SVMs) are powerful classification models, but their black box nature often limits their adoption in high-stakes domains where interpretability is crucial. The contextualized SVM framework provides a path towards explainability by training specialized models for specific data contexts, but a practical software implementation has been missing. This paper introduces XSVM@ctx-Lib as an open-source Python library that implements the contextualized SVM architecture. The library is built to be computationally efficient by parallelizing both the model training and evaluation processes. XSVM@ctx-Lib was validated against standard SVM implementations using multiple datasets, demonstrating significant performance advantages. The library achieved reductions in training and classification times. Furthermore, the contextual approach improved classification precision and accuracy by up to 9%. XSVM@ctx-Lib provides researchers and practitioners with a powerful tool that enhances SVMs by simultaneously improving transparency, computational speed, and predictive accuracy.

Automating the design of Deep Neural Networks for specific applications using Neural Architecture Search (NAS) has become increasingly attractive. However, traditional NAS methods are often hindered by the high cost of evaluating numerous candidate networks, rendering them impractical for large-scale searches. In this work, we assessed a proxy-driven Estimation of Distribution Algorithm (EDA) for optimization of topological and quantitative hyperparameters. By leveraging the ideas of designing network design spaces and iterated sampling algorithms with zero-cost proxies, our approach reduces the search cost significantly. Furthermore, the proxy driven EDA achieves near-optimal results in the NATS-Bench topology and size search spaces on CIFAR-10, CIFAR-100, and ImageNet16-120 datasets for image classification. The results show that a proxy-driven EDA can reliably explore large, mixed-type design spaces with negligible computational overhead, opening the door to rapid, resource-friendly architecture discovery.

We introduce Parallel Adaptive View Excitation (PAVE), a framework for parallel selection and training across sample-view and feature-view graph Laplacians. Each view independently decomposes the data manifold, exposing complementary spectral structure for downstream learning. Unlike unified-graph pipelines that merge all relations into a single Laplacian with inherent serial bottlenecks, PAVE exposes coarse-grained task parallelism by evaluating sample-view and feature-view manifolds concurrently and selecting the view with higher validation accuracy, while retaining the alternate view for interpretability. Across 16 benchmark datasets, PAVE matches or exceeds strong gradient-boosted baselines on most tasks, with performance degradations typically small when present, while maintaining comparable or improved F1 and AUC. On modern multi-core systems, PAVE achieves 1.5--1.9$\times$ training speedup via view-level parallelism, with approximately 60\% strong-scaling efficiency up to 32 workers and near-ideal efficiency under weak scaling. These results show that dual-view spectral excitation provides a lightweight, interpretable, and scalable alternative to conventional feature selection and manifold learning.

The Triplet Margin Ranking Loss is one of the most widely used loss functions in Siamese Networks for solving Distance Metric Learning (DML) problems. This loss function depends on a margin parameter $\mu$, which defines the minimum distance that should separate positive and negative pairs during training. In this work, we show that, during training, the effective margin of many triplets often exceeds the predefined value of $\mu$, provided that a sufficient number of triplets violating this margin is observed. This behavior indicates that fixing the margin throughout training may limit the learning process. Based on this observation, we propose a margin scheduler that adjusts the value of $\mu$ according to the proportion of easy triplets observed at each epoch, with the goal of maintaining training difficulty over time. We show that the proposed strategy leads to improved performance when compared to both a constant margin and a monotonically increasing margin scheme. Experimental results on four different datasets show consistent gains in verification performance. Our trained models and source code are available at github.com/robertotomchak/maintaining-difficulty.

Despite the empirical success of DNN, their internal training dynamics remain difficult to characterize. In ReLU-based models, the activation pattern induced by a given input determines the piecewise-linear region in which the network behaves affinely. Motivated by this geometry, we investigate whether training exhibits a two-timescale behavior: an early stage with substantial changes in activation patterns and a later stage where weight updates predominantly refine the model within largely stable activation regimes. We first prove a local stability property: outside measure-zero sets of parameters and inputs, sufficiently small parameter perturbations preserve the activation pattern of a fixed input, implying locally affine behavior within activation regions. We then empirically track per-iteration changes in weights and activation patterns across fully-connected and convolutional architectures, as well as Transformer-based models, where activation patterns are recorded in the ReLU feed-forward (MLP/FFN) submodules, using fixed validation subsets. Across the evaluated settings, activation-pattern changes decay 3 times earlier than weight-update magnitudes, showing that late-stage training often proceeds within relatively stable activation regimes. These findings provide a concrete, architecture-agnostic instrument for monitoring training dynamics and motivate further study of decoupled optimization strategies for piecewise-linear networks. For reproducibility, code and experiment configurations will be released upon acceptance.

We propose Evolutionary Dynamic Loss (EDL), a framework that learns a transferable classification loss in the probability space using unlimited synthetic prediction-label pairs, without accessing real samples during the main loss pretraining stage. EDL parameterizes the loss as a lightweight network and is trained with a semantics-free ranking-consistency objective that assigns larger penalties for more erroneous predictions. To robustly explore the space of loss functions, we optimize EDL via an evolutionary strategy and introduce chaotic mutation to improve exploration under noisy fitness evaluations. Experiments on CIFAR-10 with ResNet backbones show that EDL can serve as a drop-in replacement for cross-entropy and achieves competitive or improved accuracy, while ablation studies confirm that chaotic mutation yields faster convergence and better synthetic pretraining metrics than standard Gaussian mutation.

Transfer learning (TL) has proven highly effective at improving generalization, yet theoretical insights lag behind its rapid methodological advances. This study attempts to reduce this research gap by establishing a unified theoretical framework for analogy-based TL, grounded in a commonly used concept known as \emph{similarity}. We systematically map the three core subprocesses, i.e., retrieval, adaptation and evaluation, to the fundamental challenges in TL: what, how and when to transfer. Building upon this framework, we derive two theorems governing the gain in generalization performance of analogy-based TL: (1) the theorem of unconditionally non-negative gain, which ensures safety against negative transfer through accurate evaluation; and (2) the theorem of conditionally positive gain, contingent upon either large-scale source retrieval or effective adaptation. Furthermore, we discuss the compatibility of this framework with the No Free Lunch theorem, clarifying that the efficacy of TL arises from the alignment of analogy-based inductive biases with the non-uniform distribution of problems, rather than a violation of the theorem. These insights theoretically rationalize the paradigm shift from seeking universal transferability metrics to exploiting specialized inductive biases.

Dopaminergic signalling is central to value learning and decision making. It has been observed that multiple pathways with different patterns of connectivity project to midbrain dopaminergic neurons, some involving direct excitatory projections while others involve disinhibition. However, the respective contributions of these patterns to dopamine control, and their computational and functional advantages remain unclear. In the current work we simulate and evaluate two fully spiking neural models of dopaminergic control, based either solely on disinhibition, or solely on direct inhibitory and excitatory projections. We compare these models in terms of their engineering properties, their resulting spiking profiles, and their ability to successfully acquire representations of expected value in a 3-armed bandit task. We find that both models are able to operate at an asynchronous-irregular firing regime, but that the firing profile of the direct integration model is less resilient to disruption and more sensitive to incoming signals. In addition, the disinhibition model performs better in the learning task. We conclude that while the direct model is more parsimonious, disinhibition-based control remains advantageous in the operational context. Our results have implications for the study of decision-making brain circuits as well as for the design of brain-inspired systems.

Spiral waves are a representative class of spatiotemporal activity patterns widely observed in biological nervous systems and are closely related to neural information propagation. In this work, we investigate the emergence of spiral waves in a locally connected two-dimensional neural network composed of Aihara chaotic neurons. By systematically varying key parameters, including the local connection range, refractory-related parameters, feedback strength, network size, and boundary conditions, we examine dynamical behavior of the network under the different conditions. Numerical simulations show that spiral wave formation depends sensitively on the interplay between local connectivity and network parameters, and only emerges within specific parameter regions. Changes in network structures and parameters can significantly alter the spatiotemporal evolution of activity patterns. Our results clarify how refractoriness interact with local connectivity to regulate the emergence and stability of spiral wave patterns in neural networks.

Spatial aliasing occurs when two or more distinct locations produce highly similar place-cell representations, primarily due to environmental symmetry or repetitive structures. This issue is most pronounced when place representations are constructed solely from boundary vector cell (BVC) inputs, because symmetric or repetitive structures can yield indistinguishable sensory patterns across multiple locations in an environment. This work introduces grid cell signals to mitigate spatial aliasing in such settings. Because grid cells contribute periodic, internally generated spatial signals that vary independently of environmental geometry, they play a key role in disambiguating perceptually identical locations. We integrate multiple modules of analytically constructed grid cells with BVC-driven place cells and show that this leads to a 94--99\% reduction in spatial aliasing relative to a BVC-only baseline across three environments: an open environment without obstacles; an environment with a cross-shaped central obstacle creating high visual symmetry; and a maze environment. The greatest improvement occurs in the environment with the highest visual symmetry. These results indicate that grid cells provide information complementary to boundary-based inputs, yielding more reliable place representations in geometrically ambiguous environments.

Reward magnitudes can vary substantially from a context to another. Value coding thus requires adapting neural responses to input changes to make use of their full dynamic range. Two models have recently been proposed aiming to provide a theoretical account for context-sensitive choice valuation in the framework of reinforcement learning. One model implements range adaptation relative to the minimum and maximum achievable outcomes, building on the notions of efficient coding and signal normalization. Another model borrows concepts from intrinsic motivation and posits that valuation integrates extrinsic and intrinsic rewards. Both models are able to fit human data well, but the latter was found to perform better under specific experimental conditions. In this paper, we propose that dual range adaptation over locally and globally experienced reward ranges combines the best of the two worlds. Our model benefits from the theoretical foundations and computational advantages of range adaptation. It also displays choice behaviors qualitatively similar to the model integrating intrinsic rewards. Model fitting suggests that data previously favoring the second model are best explained by our model. Our dual range adaptation model thus offers a new mechanistic account as well as testable predictions related to the representation of value for decision-making.

Current tool-calling protocols such as the Model Context Protocol (MCP) enable AI agents to invoke external services, but their rigid transport assumptions and lack of mandatory security can be limiting in heterogeneous, event-driven IoT environments underpinning edge intelligence. This work analyzes MCP’s limitations and introduces a Web-of-Things (WoT)–based integration layer that uses standardized semantic descriptions and multiple protocol bindings to support dynamic tool discovery and subscription-based events, with security requirements declared via Thing Descriptions. We design a reference architecture around WoT and validate it using a prototype system comprising a web-based front end (implemented with Cesium map) and heterogeneous back-end tools, services, and sensor streams exposed as WoT Things. In our testbed, the prototype sustains high-frequency event interactions, restricts access to protected endpoints in automated security scanning, and completes multi-step tasks with low additional integration overhead.

Reinforcement Learning from Human Feedback (RLHF) has emerged as the dominant paradigm for aligning large language models with human preferences. However, existing approaches suffer from a fundamental limitation: collapsing multifaceted human preferences into a single scalar reward signal leads to suboptimal credit assignment and reward hacking behaviors. This work introduces Context-Aware Disentangled Reward Learning (CADRL), a novel framework that decomposes reward signals into four interpretable latent factors---accuracy, coherence, style, and safety---and dynamically weighs their importance based on task context. Through a variational autoencoder architecture constrained for interpretability, the proposed meta-reward model learns disentangled factor representations while a context encoder predicts factor-specific importance weights conditioned on prompt semantics. A hierarchical policy architecture operates at two levels: high-level factor selection guided by the meta-reward model, and low-level token generation with factor-decomposed value functions. Extensive experiments on Reddit TL;DR summarization and Anthropic HH-RLHF dialogue demonstrate that CADRL achieves 6.7\% and 6.3\% improvements over strong baselines while requiring 15.2\% less training time. The factor-level advantage estimation reduces gradient variance by 40\%, enabling more stable optimization and superior sample efficiency.

Mathematical reasoning in large language models has improved substantially with reinforcement learning using verifiable rewards, where final answers can be checked automatically and converted into reliable training signals. Most such pipelines optimize outcome correctness only, which yields sparse feedback for long, multi-step solutions and offers limited guidance on intermediate reasoning errors. Recent work therefore introduces process reward models (PRMs) to score intermediate steps and provide denser supervision. In practice, PRM scores are often imperfectly aligned with final correctness and can reward locally fluent reasoning that still ends in an incorrect answer. When optimized as absolute rewards, such signals can amplify fluent failure modes, induce reward hacking, and destabilize policy updates. Existing PRM-based methods improve PRM quality, filter trajectories, or modify reward aggregation, but they do not directly constrain how process rewards interact with outcome correctness during optimization. We propose PROGRS, a framework that leverages PRMs while keeping outcome correctness dominant. PROGRS treats process rewards as relative preferences within outcome groups rather than absolute targets. We introduce outcome-conditioned centering, which shifts PRM scores of incorrect trajectories to have zero mean within each prompt group. It is removing systematic bias while preserving informative rankings. To stabilize process signals, PROGRS combines a frozen quantile-regression PRM with a multi-scale coherence evaluator that penalizes short-window confidence volatility. We integrate the resulting centered process bonus into Group Relative Policy Optimization (GRPO) without auxiliary objectives or additional trainable components. Across MATH-500, AMC, AIME, MinervaMath, and OlympiadBench, PROGRS consistently improves Pass@1 over outcome-only baselines (e.g., 74.9% vs.69.7% on MATH-500; 59.0% vs.52.0% on AMC-2023) and achieves stronger performance with fewer rollouts. These results show that outcome-conditioned centering enables safe and effective use of process rewards for mathematical reasoning.

Using large language models (LLMs) to improve generalization across unseen out-of-domain tasks is promising in neural architecture search (NAS), yet it still struggles to solve regression problems precisely when performance distributions shift across domains. To address the issue, the proposed framework strategically integrates several existing methods: (1) employing rank-based contrastive loss, which focuses on learning-to-rank by the candidates relative comparisons, (2) take advantage of hierarchical pairwise sampling to emphasize fine-grained distinctions among top-performing architectures but still maintain a global ordering and (3) utilizing decoder-based large language models to enable potential inference ability gains. To verify the performance of the framework, it is evaluated on the NAS-Bench-201 and TransNAS-Bench-101 benchmarks. Experimental results provide strong evidence that the proposed strategic integration enables LLMs to adapt to regression problems and demonstrates significantly more effective transferability, achieving superior zero-shot performance compared to conventional strong predictors.

Active learning (AL) reduces labeling cost by querying informative samples, but in tabular settings its cold-start gains are often limited because uncertainty estimates are unreliable when models are trained on very few labels. Tabular foundation models such as TabPFN provide calibrated probabilistic predictions via in-context learning (ICL), i.e., without task-specific weight updates, enabling an AL regime in which the labeled context - rather than parameters - is iteratively optimized. We formalize Tabular Active In-Context Learning (Tab-AICL) and instantiate it with four acquisition rules: uncertainty (TabPFN-Margin), diversity (TabPFN-Coreset), an uncertainty-diversity hybrid (TabPFN-Hybrid), and a scalable two-stage method (TabPFN-Proxy-Hybrid) that shortlists candidates using a lightweight linear proxy before TabPFN-based selection. Across 20 classification benchmarks, Tab-AICL improves cold-start sample efficiency over retrained gradient-boosting baselines (CatBoost-Margin and XGBoost-Margin), measured by normalized AULC up to 100 labeled samples.

Reinforcement learning (RL) agents often struggle with out-of-distribution (OOD) scenarios, leading to high uncertainty and random behavior. While language models (LMs) contain valuable world knowledge, larger ones incur high computational costs, hindering real-time use, and exhibit limitations in autonomous planning. We introduce Adaptive Safety through Knowledge (ASK), which combines smaller LMs with trained RL policies to enhance OOD generalization without retraining. ASK employs Monte Carlo Dropout to assess uncertainty and queries the LM for action suggestions only when uncertainty exceeds a set threshold. This selective use preserves the efficiency of existing policies while leveraging the language model’s reasoning in uncertain situations. In experiments on the FrozenLake environment, ASK shows no improvement in-domain, but demonstrates robust navigation in transfer tasks, achieving a reward of 0.95. Our findings indicate that effective neuro-symbolic integration requires careful orchestration rather than simple combination, highlighting the need for sufficient model scale and effective hybridization mechanisms for successful OOD generalization.

Complex-valued neural networks (CVNNs) are rising in popularity for all kinds of applications. To safely use CVNNs in practice, analyzing their robustness against outliers is crucial. One well known technique to understand the behavior of deep neural networks is to investigate their behavior under adversarial attacks, which can be seen as worst case minimal perturbations. We design Phase Attacks, a kind of attack specifically targeting the phase information of complex-valued inputs. Additionally, we derive complex-valued versions of commonly used adversarial attacks. We show that in some scenarios CVNNs are more robust than RVNNs and that both are very susceptible to phase changes with the Phase Attacks decreasing the model performance more, than equally strong regular attacks, which can attack both phase and magnitude.

Subtask distillation is an emerging paradigm in which compact, specialized models are extracted from large, general-purpose 'foundation models' for deployment in environments with limited resources or in standalone computer systems. Although distillation uses a teacher model, it still relies on a dataset that is often limited in size and may lack representativeness or exhibit spurious correlations. In this paper, we evaluate established distillation methods, as well as the recent SubDistill method, when using data with spurious correlations for distillation. As the strength of the correlations increases, we observe a widening gap between advanced methods, such as SubDistill, which remain fairly robust, and some baseline methods, which degrade to near-random performance. Overall, our study underscores the challenges of knowledge distillation when applied to imperfect, real-world datasets, particularly those with spurious correlations.

Incremental learners deployed on streaming data must remain robust to evolving adversarial perturbations, yet most adversarial-robustness studies assume offline multi-epoch training with repeated access to historical data. We investigate adversarial robustness in Fuzzy ARTMAP, a prototype-based Adaptive Resonance Theory (ART) model that supports single-pass learning without replay. To enable strong adaptive evaluation despite ARTMAP’s non-differentiable winner-take-all dynamics, we propose WB-Softmax, a differentiable softmax relaxation that aggregates category-level activations into class-level scores for gradient-based attacks. Across USPS, MNIST, and Fashion-MNIST, WB-Softmax Projected Gradient Descent (PGD) achieves 89–100% attack success on vanilla models, exceeding transfer and query-based baselines at matched budgets and satisfying anti-gradient-masking sanity checks. We then study adversarial training under true streaming constraints by comparing offline versus online adversarial example generation and standard versus selective updates. Offline adversarial training consistently collapses robustness despite substantial category growth, whereas online training is dataset dependent. We introduce progressive two-stage selective training that combines selective filtering with ϵ scheduling, achieving the best overall robustness (AURAC) on all three datasets (28.2%, 64.5%, and 41.3%). Finally, leveraging ART’s explicit category geometry, we monitor incremental cluster validity indices to diagnose separation collapse and design a separation-aware absorption rule that improves high-ϵ robustness on USPS (11.0±1.4% at ϵ=0.30).

Adversarial examples exploit small perturbations in the input data to manipulate the decisions of neural network classifiers. To address this challenge, verification tools have been developed to certify a range of perturbations that cannot alter the prediction of a classifier. Nevertheless, classifiers can be implemented using different precisions and arithmetic due to computational or hardware constraints. Recent studies reveal that discrepancies between the implemented and the verified classifier can be exploited to invalidate existing guarantees, since verifiers predominantly rely on floating-point arithmetic. In this work, we show that state-of-the-art verifiers (i.e., the podium of the VNN-COMP) remain vulnerable to this class of attacks by introducing a new, computationally efficient, precision-based evasion attack. We then investigate how augmenting a verifier with interval-based arithmetic can effectively mitigate these attacks, thereby enabling precision-aware verification. As our empirical evaluation confirms, our approach ensures consistent guarantees and successful mitigation.

Adversarial attacks against deep neural networks are commonly constructed under Lp norm constraints, most often using p = 1, p = 2 or p = infinity, and potentially regularized for specific demands such as sparsity or smoothness. These choices are typically made without a systematic investigation of how the norm parameter p influences the structural and perceptual properties of adversarial perturbations. In this work, we study how the choice of p affects sparsity and smoothness of adversarial attacks generated under Lp norm constraints for values of p in [1, 2]. To enable a quantitative analysis, we adopt two established sparsity measures from the literature and introduce three smoothness measures. In particular, we propose a general framework for deriving smoothness measures based on smoothing operations and additionally introduce a smoothness measure based on first-order Taylor approximations. Using these measures, we conduct a comprehensive empirical evaluation across multiple real-world image datasets and a diverse set of model architectures, including both convolutional and transformer-based networks. We show that the choice of L1 or L2 is suboptimal in most cases and the optimal p value is dependent on the specific task. In our experiments, using Lp norms with p in [1.3, 1.5] yields the best trade-off between sparse and smooth attacks. These findings highlight the importance of principled norm selection when designing and evaluating adversarial attacks.

Attacking semantic segmentation models is significantly harder than image classification models because an attacker must flip thousands of pixel predictions simultaneously. Standard pixel-wise cross-entropy (CE) is ill-suited to this setting: it tends to overemphasize already-misclassified pixels, which slows optimization and overstates model robustness. To address these issues, we introduce TsallisPGD, an adversarial attack built on the Tsallis cross-entropy, a generalization of CE parameterized by $q$, which adaptively reshapes the gradient landscape by controlling gradient concentration across pixels. By varying $q$, we steer the attack toward pixels at different confidence levels. We first show that no single fixed-$q$ is universally optimal, as its effectiveness depends on the dataset, model architecture, and perturbation budget. Motivated by this, we propose a dynamic $q$-schedule that sweeps $q$ during optimization. Extensive experiments on Cityscapes, Pascal VOC, and ADE20K show that TsallisPGD, using a single validation-selected schedule, achieves the best average attack rank across all evaluated settings and improves over CEPGD, SegPGD, CosPGD, JSPGD, and MaskedPGD in reducing accuracy and mIoU on both standard and robust models.

Proteins are central to virtually all biological processes, and their structural diversity underlies a wide range of functions. Among them, multifunctional proteins (i.e., those associated with multiple enzymatic activities) pose unique challenges for computational analysis. In this work, we apply a suite of machine learning and deep learning techniques to predict multifunctionality roles starting from protein 3D structures. Proteins are represented as residue contact networks, enabling the use of graph machine learning approaches. We explore feature-based methods grounded in simplicial complexes analysis alongside end-to-end Graph Neural Networks implementing recent message-passing schemes. Our models address the task of multi-label classification of the first level Enzyme Commission numbers of multifunctional proteins. Evaluation across a strictly multifunctional subset of the human proteome associated with repeated stratified validation demonstrates that structural graph representations effectively capture signals of multifunctionality, shedding light on how protein architecture encodes a diverse range of biochemical roles.

Neutrino oscillations encode fundamental information about neutrino masses and mixing parameters, offering a unique window into physics beyond the Standard Model. Estimating these parameters from oscillation probability maps is, however, computationally challenging due to the maps’ high dimensionality and nonlinear dependence on the underlying physics. Traditional inference methods, such as likelihood-based or Monte Carlo sampling approaches, require extensive simulations to explore the parameter space, creating major bottlenecks for large-scale analyses. In this work, we introduce a data-driven framework that reformulates atmospheric neutrino oscillation parameter inference as a supervised regression task over structured oscillation maps. We propose a hierarchical transformer architecture that explicitly models the two-dimensional structure of these maps, capturing angular dependencies at fixed energies and global correlations across the energy spectrum. To improve physical consistency, the model is trained using a surrogate simulation constraint that enforces agreement between the predicted parameters and the reconstructed oscillation patterns. Furthermore, we introduce a neural network-based uncertainty quantification mechanism that produces distribution-free prediction intervals with formal coverage guarantees. Experiments on simulated oscillation maps under Earth-matter conditions demonstrate that the proposed method is comparable to a Markov Chain Monte Carlo baseline in estimation accuracy, with substantial improvements in computational cost (around 240$\times$ fewer FLOPs and 33$\times$ faster in average processing time). Moreover, the conformally calibrated prediction intervals remain narrow while achieving the target nominal coverage of 90\%, confirming both the reliability and efficiency of our approach.

Astronomical brokers must filter millions of alerts to prioritize follow-up in surveys such as the Zwicky Transient Facility (ZTF) and the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST). To address this challenge we present Score-based Log-Density Network (SLDNet), an unsupervised generative model that estimates the gradient of the data log-density in a multiscale manner. Trained with denoising score matching, SLDNet defines an anomaly score using a conservative potential as a proxy for the negative log-density, avoiding the calculation of the normalization constant.

We evaluate SLDNet on two complementary protocols: (i) leave-one-class-out on 152 tabular features extracted by the Automatic Learning for the Rapid Classification of Events (ALeRCE) broker for ZTF, where SLDNet achieves the best area under the receiver operating characteristic curve in most transient and stochastic subclasses; and (ii) Out-of-distribution discovery and Rare-in-distribution scenarios following the Distance Multi-Metric Anomaly Detection (DiMMAD) evaluation protocol on the Extended LSST Astronomical Time-series Classification Challenge (ELAsTiCC) and ZTF using Supernova Parametric Model features, reporting area under the precision–recall curve and purity/diversity vs. follow-up budget. To our knowledge, this is the first study validating multiscale log-density estimation based on denoising score matching for anomaly detection in the astronomical domain.

Hyperspectral image classification is often challenged by severe class imbalance and limited labeled samples, which significantly degrade the performance of conventional supervised learning methods. To address these issues, this paper proposes a semi-supervised learning framework that iteratively incorporates informative unlabeled samples based on neighborhood consistency to improve minority-class representation. Unlike synthetic oversampling techniques such as SMOTE, the proposed method leverages the intrinsic data distribution without generating artificial samples.

Extensive experiments under varying class imbalance ratios demonstrate that the proposed approach consistently outperforms baseline preprocessing methods in terms of Recall, F1-score, G-mean, and AUC. Paired t-tests further confirm that the improvements over SMOTE are statistically significant, particularly under severe imbalance conditions. Moreover, the proposed framework achieves stable performance across different base classifiers and exhibits low sensitivity to hyperparameter settings, highlighting its robustness and practical applicability.

Audio-to-image retrieval offers an interpretable alternative to audio-only classification for bioacoustic species recognition, but learning aligned audio–image representations is challenging due to the scarcity of paired audio–image data. We propose a simple and data-efficient approach that enables audio-to-image retrieval without any audio–image supervision. Our proposed method uses text as a semantic intermediary: we distill the text embedding space of a pretrained image–text model (BioCLIP-2), which encodes rich visual and taxonomic structure, into a pretrained audio–text model (BioLingual) by fine-tuning its audio encoder with a contrastive objective. This distillation transfers visually grounded semantics into the audio representation, inducing emergent alignment between audio and image embeddings without using images during training. We evaluate the resulting model on multiple bioacoustic benchmarks. The distilled audio encoder preserves audio discriminative power while substantially improving audio–text alignment on focal recordings and soundscape datasets. Most importantly, on the SSW60 benchmark, the proposed approach achieves strong audio-to-image retrieval performance exceeding baselines based on zero-shot model combinations or learned mappings between text embeddings, despite not training on paired audio–image data. These results demonstrate that indirect semantic transfer through text is sufficient to induce meaningful audio–image alignment, providing a practical solution for visually grounded species recognition in data-scarce bioacoustic settings.

Reliable reconstruction of urban flooding is important for urban planning, flood monitoring, and risk analysis; however, progress is hindered by the limited availability of datasets with usable scene geometry and flood-depth annotations. Existing approaches often rely on LiDAR scanning, 3D reconstruction, or purely generative models, which can be computationally expensive and may not enforce physically realistic fluid dynamics over time. Moreover, obtaining labeled ground truth data —particularly flood depth —is costly and difficult to scale. To address these limitations, we propose a monocularvideo- driven framework that combines open-vocabulary semantic segmentation, monocular depth estimation, and a lightweight physics-based simulation model using Shallow Water Equations (SWE). The simulation operates on scene-derived terrain and dynamic obstacles to produce temporally consistent flooding evolution, which is then used to synthesize flooding imagery and corresponding dense supervision signals. As a proof of concept, we also demonstrate a physics-informed neural surrogate model that regularizes predictions with SWE residuals. Overall, the proposed framework enables efficient generation of physically grounded synthetic data for training and benchmarking flood segmentation and depth estimation models, supporting data-driven urban resilience and disaster preparedness.

Accurate residential indoor room segmentation is crucial for mobile robots, but traditional geometric methods often fail in cluttered environments, while vision-based approaches are still dependent on geometric analysis, and therefore limited to simple environments with clear geometric boundaries. We propose a framework that can handle complex room boundaries by using a robot's visual input to simultaneously segment and classify rooms. This is achieved by projecting image-based room predictions onto a 2D occupancy map using spatial aggregation techniques. By leveraging the vision cone of the robot and powerful image classification models such as CNNs, ViTs and VLMs, the map is segmented into rooms with a higher Intersection over Union (IoU) than traditional methods while handling environments with complex geometry. The results demonstrate that our method using VLMs and vision cone aggregation outperforms other approaches, achieving up to 76.22\% classification accuracy on ArchitectThor room segmentation without any task-specific training, offering superior accuracy and flexibility for robotic semantic mapping, while overcoming the geometry-dependency of traditional methods. We also provide the code and improved dataset.

Recent advances in robot learning increasingly rely on LLM-based task planning, leveraging their ability to bridge natural language with executable actions. While prior works showcased great performances, the widespread adoption of these models in robotics has been challenging as 1) existing methods are often closed-source or computationally intensive, neglecting the actual deployment on real-world physical systems, and 2) there is no universally accepted, plug-and-play representation for robotic task generation. Addressing these challenges, we propose BTGenBot-2, a 1B-parameter open-source small language model that directly converts natural language task descriptions and a list of robot action primitives into executable behavior trees in XML. Unlike prior approaches, BTGenBot-2 enables zero-shot BT generation, error recovery at inference and runtime, while remaining lightweight enough for resource-constrained robots. We further introduce the first standardized benchmark for LLM-based BT generation, covering 52 navigation and manipulation tasks in NVIDIA Isaac Sim. Extensive evaluations demonstrate that BTGenBot-2 consistently outperforms GPT-5, Claude Opus 4.1, and larger open-source models across both functional and non-functional metrics, achieving average success rates of 90.38% in zero-shot and 98.07% in one-shot, while delivering up to 16× faster inference compared to the previous BTGenBot.

As Vision-Language-Action (VLA) models are increasingly used in complex, multi-task robotic setups, automatic detection of execution failures has become critical for safety and reliability. Recent studies have demonstrated that latent action embeddings of a VLA can be leveraged to predict task success. However, relying solely on action embeddings neglects the environmental context that often precipitates failure. In this paper, we address this gap by first demonstrating that the distributions of image embeddings correlate with failure scenarios. Motivated by this finding, we propose SAFER, a fusion approach that integrates both action and image embeddings into a dedicated failure prediction module. Through extensive experiments, we show that image embeddings provide complementary information to action embeddings, providing up to 8% point improvement in seen and 2% point improvement on unseen tasks with 5.7x times faster inference over the state-of-the-art for detecting failures. Overall, SAFER underscores the necessity of visual context in monitoring the internal reasoning of VLA models and provides a more robust framework for autonomous failure detection. Code and trained models are available at: https://github.com/purentap/safer

Curriculum reinforcement learning (CRL) aims to speed up agent learning by organizing tasks in a progressively harder sequence. However, many existing CRL methods struggle to guide agents toward meaningful goals, especially when domain knowledge is limited or unavailable. To overcome this limitation, we introduce PIDCRL, a diffusion-based CRL framework that leverages a mask-conditional, image-to-image pretrained diffusion model to automatically generate curriculum goals from agent trajectory heatmaps. Given a heatmap trajectory image as input, the model produces candidate curriculum goals, which we then filter and select using several strategies, including averaging, Q- value scoring, and a trainable reward-based mechanism. These strategies identify curriculum goals that are both achievable and appropriately challenging, enabling effective curricula without expert-designed heuristics. We validate PIDCRL across three maze environments, where it matches or outperforms ten state-of-the-art CRL baselines.

Large Language Models (LLMs) are increasingly used in autonomous robotics, often in end-to-end frameworks that map natural language instructions directly to actions. While effective in constrained settings, such approaches suffer from limited grounding, poor explainability, and weak knowledge reuse. We propose an alternative integration strategy in which LLMs act as a high-level cultural knowledge repository that informs, but does not control, a robotic cognitive architecture. All perception, learning, and action selection remain grounded in embodied interaction. We show that leveraging LLM-provided priors during exploration in previously unknown domains significantly accelerates the acquisition of grounded goals compared to intrinsically motivated exploration. Experiments in simulation and on a real robot demonstrate improved early learning efficiency, goal discovery, and interpretability. Our results suggest that embedding LLMs within hybrid cognitive architectures enables scalable, explainable, and open-ended robotic learning.

Reinforcement learning (RL) often struggles in real-world tasks with high-dimensional state spaces and long horizons, where sparse or fixed rewards severely slow down exploration and cause agents to get trapped in local optima. This paper presents a fuzzy-logic-based reward shaping method that integrates human intuition into RL reward design. By encoding expert knowledge into adaptive, interpretable terms, fuzzy rules promote stable learning and reduce sensitivity to hyperparameters. The proposed method leverages these properties to adapt reward contributions based on the agent’s state, enabling smoother transitions between fast motion and precise control in challenging navigation tasks. The extensive simulation results on autonomous drone racing benchmarks show stable learning behavior and consistent task performance across scenarios of increasing difficulty. The proposed method achieves faster convergence and reduced performance variability across training seeds in more challenging environments, with success rates improving by up to approximately 5% compared to non-fuzzy reward formulations.

Dense prediction requires representations that are multi-scale, semantically aligned, and resolution stable, yet existing hierarchical vision transformers are task-specific, optimization-sensitive, and inefficient under scale variation. We propose MAHViT, a self-supervised hierarchical vision transformer designed as a multi-scale dense feature extractor. MAHViT introduces a dual-Layer Normalization data-flow that stabilizes deep hierarchical attention by jointly preserving gradient flow and feature diversity, enabling scalable self-supervised pre-training. To efficiently connect hierarchical encoders with dense decodes, we propose a Feature Diversity Sampling (FDS) neck that suppresses redundant tokens while retaining semantically distinct features. Pre-trained using a BYOL-based objective with resolution-adaptive alignment, MAHViT encoder learns scale-consistent representations. The encoder is evaluated on semantic segmentation task by pairing a Progressive Upsampling decoder with the FDS Neck. Through extensive evaluations, MAHViT achieves state-of-the-art or competitive multi-scale mIoU on COCO-2017, ADE20K, and Cityscapes across multiple model scales, while maintaining favorable FLOPs–accuracy trade-offs suitable for edge deployment.

Air pollution, particularly particulate matter (PM) concentration, poses a serious threat to public health and the environment. Accurate short-term forecasting of PM levels based on historical concentration data is essential for timely decision-making, public warning systems, and effective air quality management. In this study, we propose an interpretable and data-lightweight approach for predicting PM2.5 and PM10 concentrations using a model that combines temporal attention with a Kolmogorov-Arnold Network. The proposed model is evaluated using real-world air quality data collected without using meteorological or other auxiliary data. Its performance is assessed using standard error metrics, including mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). The results demonstrate that the proposed model consistently outperforms state-of-the-art deep learning and statistical methods.

Fake reviews undermine the credibility of online platforms, yet most existing detection methods rely on manually designed pipelines with extensive feature engineering and heuristic model selection, limiting robustness and cross-domain generalization. To address these challenges, this paper proposes a unified AutoML-based framework for fake review detection that automates model selection, feature optimization, and hyperparameter tuning in an end-to-end manner, evaluated in both text-only and multimodal settings. The framework is evaluated across multiple AutoML systems and five benchmark datasets, including Amazon, FRD, Yelp-NYC, TripAdvisor, and a synthetic E-Commerce dataset, and integrates textual, numerical, and visual information during the multimodal stage. Experimental results show that multimodal AutoML consistently outperforms traditional machine learning models and remains competitive with state-of-the-art deep learning approaches, achieving up to 10% accuracy increment and 9.1% F1-score, demonstrating AutoML’s effectiveness as a scalable and deployment-ready solution.

Embedding-based retrieval with Vision-Language Models (VLMs) enables high-performance text-image retrieval by projecting queries and candidates into a shared vector space and ranking them via cosine similarity. However, setting relevance thresholds manually requires effort and lacks adaptability to differences across users, queries, and models. To overcome this, we introduce a fast, lightweight score calibration method that normalizes cosine similarity scores using a formula based on the mean of estimated non-relevant candidates. This approach lessens sensitivity to candidate set size and relevance ratio. By interpreting the calibrated scores as relevance probabilities, we can estimate precision and recall at any threshold, making it possible for users to intuitively specify preferred metrics and directly control results. Experiments on diverse datasets show that our calibration method improves text-image relevance classification at global thresholds and delivers more stable retrieval quality according to user-specified criteria, allowing intuitive and consistent control over the relation between threshold and retrieval performance.

Benchmarking is essential for the evaluation of evolutionary multi-objective optimization (EMO) algorithms, yet benchmarking results are sensitive to experimental settings such as test problems, compared algorithms, performance indicators, and their parameter specifications. Recently, large language models (LLMs) have been increasingly used in EMO research and academic activities (such as assisting with paper reviews). As a result, LLMs may influence how benchmarking studies are conducted, especially for newcomers to the EMO field. This position paper examines the reliability of benchmarking suggestions generated by two LLMs by comparing their responses with those reported in earlier studies at IEEE CEC 2025. Using the same prompts and experimental settings, we analyze similarities and differences in the suggested algorithms, test problems, and performance indicators for benchmarking EMO algorithms between the two LLMs and also between the two years. In addition, we report benchmarking suggestions from the LLMs for constrained multi-objective optimization. The observations indicate that, while many LLM-generated suggestions remain consistent with widely used benchmarking practices, variations and inaccuracies are also observed. These findings highlight the need for careful interpretation of LLM-generated suggestions and emphasize the importance of benchmarking guidelines when evaluating EMO algorithms.

Neuromorphic Computing (NC) and Evolutionary Computation (EC) are mature, yet largely fragmented. This position paper argues that their convergence is timely and technically plausible. We define Neuromorphic Evolutionary Computation (NEC) as a paradigm that bridges NC and EC. We review the current landscape and its limitations, with an emphasis on the prevalence of hybrid execution and incomplete energy-latency reporting. We then present a probable Cortico-Basal-Thalamic Loop (CBTL)-inspired conceptual architecture for operator selection and outline what must be implemented on-chip to achieve an NEC approach, including operator embodiment, associative memory, and state estimation. We summarise open challenges spanning numerical precision, scalability, learning rules, benchmarking, theory, and tooling, including the need for convergence analysis framed in terms of the best-so-far process. We close with a phased roadmap and concrete reporting requirements towards a reproducible NEC ecosystem.

Identifying high-quality feature subsets for decision-makers by wrapper-based multi-objective feature selection (MOFS) has been attracting increasing attention. As a mainstream approach, evolutionary methods offer distinct advantages but also struggle with increasingly complex application scenarios and high-dimensional data, mainly in terms of time efficiency, exponentially expanding search space, and adaptive determination of a suitable classifier. To overcome these challenges, this paper proposes two simple yet effective methods: Fast Initialization (FI) and one-generation Adaptive K-Nearest Neighbor (AK). FI leverages mutual information and tournament selection to locate high-quality initial feature subsets computationally efficiently. AK verifies that, with theoretical proof, using a single generation can determine the most suitable KNN for different data to improve feature selection performance with very little time overhead and without any data analysis or assumption. Experiments on 20 real-world high-dimensional datasets demonstrate the superior performance of FI and AK to advanced initialization and KNN methods for MOFS. We also validated that the obtained feature subsets generalize well to an LLM for tabular data, enabling it to be seamlessly applied to high-dimensional data and achieve superior performance.

Artificial neural networks (ANNs), as core components of modern deep learning (DL) systems, lack the adaptive flexibility and long-term stability exhibited by biological systems. This limitation largely stems from the fact that conventional ANNs rely on uniform, local, and gradient-based parameter updates, while neglecting internal learning principles that are biological mechanisms such as neurotransmitters signalling or neuroplasticity. Consequently, many existing approaches focus on architectural expansion or mathematical fine-tuning techniques such as regularisation or parameter isolation. Inspired by the superior adaptability and plasticity of mammalian brains, we posit that neuromodulation with neurotransmitters constitutes a third axis of learning, complementary to neural activity and synaptic plasticity, and should be explicitly modelled in artificial neural networks. In this positional paper, we argue that incorporating neuromodulatory principles into ANN design represents a promising and underexplored research direction, and we advocate for greater attention to this perspective in the development of adaptive and continual learning systems.

This work proposes two necessary and sufficient conditions for neural computing namely, first, a Hilbert space H and, second, non-linear transformations in H. The conventional deep learning (neural computing) paradigm is extended hierarchically from the Euclidean (Hilbert) space R^N to the Hilbert space G_1^N, where G_1 is a quotient of equivalence classes of Generalized Intervals' Numbers (GINs). The convex cone F_1 of G_1 is considered, where an element of F_1, namely Intervals' Number (IN), is an (information) granule that may represent a probability/possibility distribution including real numbers. It turns out that F_1 is a sublattice of lattice G_1, where partial order represents semantics; moreover, axiomatic logic can be applied in F_1. Because all the required mathematical instruments are available, neural computing can be pursued in F_1^N including logic/reasoning toward eXplainable Artificial Intelligence (XAI). The far-reaching potential of proposed tools in practical applications is discussed.

The potential of AI researchers in scientific discovery remains largely untapped. Over the past decade, AI for Science (AI4Science) publications in 145 Nature Index journals have increased fifteen-fold, yet they still account for less than 3% of the total publications. Drawing upon the Diffusion of Innovation theory, we project AI4Science’s share of total publications to rise from 2.72% in 2024 to approximately 20% by 2050. Achieving this shift requires fully harnessing the potential of AI researchers, as nearly 95% of AI-driven research in these journals is led by experimental scientists. To facilitate this, we propose structured workflows and strategic interventions to position AI researchers at the forefront of scientific discovery. Specifically, we identify three critical pathways: equipping experimental scientists with accessible AI tools to amplify the impact of AI researchers, bridging cognitive and methodological gaps to enable more direct involvement in scientific discovery, and proactively fostering a thriving AI-driven scientific ecosystem. By addressing these challenges, we aim to empower AI researchers as key drivers of future scientific breakthroughs.

Brain-computer interfaces (BCIs) enable direct communication between the brain and external devices, providing critical support for individuals with motor impairments. However, accurate motor imagery (MI) decoding from electroencephalography (EEG) remains challenging due to noise and cross-session variability. This study introduces EEG-MFTNet, a novel deep learning model based on the EEGNet architecture, enhanced with multi-scale temporal convolutions and a Transformer encoder stream. These components are designed to capture both short and long-range temporal dependencies in EEG signals. The model is evaluated on the SHU dataset using a subject-dependent cross-session setup, outperforming baseline models, including EEGNet and its recent derivatives. EEG-MFTNet achieves an average classification accuracy of 58.9% while maintaining low computational complexity and inference latency. The results highlight the model’s potential for real-time BCI applications and underscore the importance of architectural innovations in improving MI decoding. This work contributes to the development of more robust and adaptive BCI systems, with implications for assistive technologies and neurorehabilitation.

Diffusion models trained on noisy datasets often reproduce high-frequency training artifacts, significantly degrading generation quality. To address this, we propose SCoRe (Spectral Cutoff Regeneration), a training-free, generation-time spectral regeneration method for clean image generation from diffusion models trained on noisy images. Leveraging the spectral bias of diffusion models, which infer high-frequency details from low-frequency cues, SCoRe suppresses corrupted high-frequency components of a generated image via a frequency cutoff and regenerates them via SDEdit. Crucially, we derive a theoretical mapping between the cutoff frequency and the SDEdit initialization timestep based on Radially Averaged Power Spectral Density (RAPSD), which prevents excessive noise injection during regeneration. Experiments on synthetic (CIFAR-10) and real-world (SIDD) noisy datasets demonstrate that SCoRe substantially outperforms post-processing and noise-robust baselines, restoring samples closer to clean image distributions without any retraining or fine-tuning.

The online information ecosystem is shaped by collaborative and adaptive processes in which narratives evolve through collective interpretation and reframing. Anticipating these dynamics is essential for developing trustworthy and human-centered AI systems that can provide proactive, interpretable support to human decision-makers before problematic narratives reach critical diffusion. However, most existing approaches rely on post-hoc analyses, offering limited utility for early intervention and pre-bunking strategies. This paper presents a collaborative multi-agent simulation framework based on Large Language Models (LLMs) to study the evolution of news narratives. Heterogeneous agents interact within a controlled network to iteratively generate and reframe news headlines, producing structured trajectories of narrative transformation. A notion of simulated Narrative Impact to quantify a news item’s capacity to organize, sustain, and attract collective narrative activity over time, is formulated. Building on this formulation, an early alerting task predicts the structural impact of a narrative from partial observations of the simulation, enabling the identification of high-impact narratives at early stages. Experimental results, obtained using an XGBoost-based predictive model, demonstrate that the Narrative Impact can be accurately identified (R2 = 0.71) at early stages, even in the absence of intentional misinformation.

Federated Learning (FL) enables collaborative model training across distributed nodes while preserving data privacy. However, its decentralized architecture introduces significant security vulnerabilities, particularly susceptibility to poisoning attacks where malicious participants inject misleading information into the training process. In this paper, we propose a Multi-perspective Poisoning Detection approach that based on a Graph Convolutional Network architecture that integrates both temporal and spatial perspectives within a unified framework. This integration enables more robust and comprehensive drift analysis by applying simultaneous spatio-temporal evaluation to all detected drift instances. We evaluate the proposed approach on a new dataset generated starting from a public dataset. Experimental results demonstrate that the integrated hierarchical approach achieves superior detection accuracy with respect to the plain transformer-based autoencoder baseline and reduced false positive rates, enhancing the resilience of federated learning systems against sophisticated poisoning attacks while maintaining adaptability to legitimate concept drift.

Federated learning in resource-constrained environ-ments poses a complex multi-objective optimization problemcharacterized by extreme data heterogeneity, stochastic client availability, and fluctuating energy constraints. Traditional clientselection mechanisms, relying predominantly on static heuristicsor single-objective optimization, lack the flexibility to navigate thedynamic Pareto frontier between data utility, system robustness,and energy efficiency. We propose FedE2A, a framework that reframes client selection as a context-aware decision process governed by a Large Language Model (LLM). In this architecture, clients are abstracted into structured Quality Profile Vectors (QPV) that encapsulate strictly local states—including epistemic uncertainty via Monte Carlo Dropout, system health, and historical reliability. The server-side LLM acts as a reasoning engine, interpreting these high-dimensional profiles to dynamically arameterize the selection policy based on the current training phase and convergence trajectory. Unlike rigid rule based systems, this neuro-symbolic approach allows for semantic interpretation of system anomalies and adaptive balancing of competing objectives. Extensive experiments on CIFAR-10 under non-IID settings and label noise demonstrate that FedE2A yields significant improvements in convergence stability and energy efficiency compared to state-of-the-art baselines, establishing LLM-based governance as a principled methodology for robust orchestration in heterogeneous federated systems.

Regression-based stability prediction is a core component of decentralized smart grid control, yet its security under adversarial data-injection attacks remains largely unexplored. Unlike prior work that focuses on binary stability classification, many operational systems rely on continuous stability scores that directly influence control decisions. In this paper, we study black-box adversarial attacks against regression-based stability prediction under realistic query and injection budget constraints. We adapt and evaluate multiple query-based derivative-free attacks designed to directly maximize regression output distortion rather than misclassification. We further introduce an uncertainty-aware sampling strategy that allows an attacker to concentrate a limited injection budget on the most vulnerable inputs, significantly amplifying attack impact.

To mitigate these attacks, we propose Stochastic Quantile Smoothing (SQS), a lightweight, query-side defense that randomizes model outputs using secret quantile sampling and input smoothing. SQS returns a single stochastic but plausible scalar per query, deliberately obstructing gradient and expectation-based black-box optimization while preserving utility for legitimate users. Extensive experiments on smart grid stability prediction show that SQS substantially reduces the effectiveness of black-box attacks by increasing adversarial query cost, while incurring only modest degradation in predictive performance.

The ever increasing adoption of residential photovoltaic (PV) systems fundamentally alters true household consumption and therefore challenges the applicability of conventional Non-Intrusive Load Monitoring (NILM) techniques. In this paper, we propose a physics-guided neural network for the disaggregation of household appliances, as well as behind the meter solar generation using the net load signal and local irradiance measurements. Unlike conventional methods that regress solar power directly, our method introduces a latent Correction Factor (CF) that is estimated alongside appliance power profiles. This predicted CF scales the irradiance input to reconstruct solar generation, strictly enforcing physical constraints (e.g. zero output under no irradiance) directly into the learning process. This formulation enables the CF to implicitly capture PV system efficiency variations due to environmental and operational factors without explicit supervision. Experimental evaluation on real world datasets demonstrate that the proposed approach achieves improved disaggregation accuracy and enhanced consistency compared to unconstrained solar regression baselines in effectively decoupling household consumption from distributed generation in PV-integrated environments.

The rapid growth of electric vehicle charging loads is creating new operational stress in power distribution systems, where additional demand can trigger voltage violations and equipment overloads. To manage this risk, utilities use hosting capacity analysis to assess how much new charging load can be accommodated at each location without violating operating limits. Existing learning-based estimators can provide fast predictions, but many implicitly assume a fixed network configuration and therefore their accuracy degrades when switching operations alter feeder topology during maintenance, restoration, or load balancing. This paper introduces a topology-aware spatiotemporal learning framework that predicts time-varying hosting capacity at the bus level while remaining robust to feeder reconfiguration. The method represents the feeder as a graph with physically meaningful edge descriptors and performs joint spatial-temporal reasoning over a short history window using an edge-conditioned attention mechanism. Experiments on a real-world, smart-meter-informed distribution system under multiple reconfiguration scenarios show that the proposed approach maintains high accuracy on unseen topologies and consistently outperforms topology-agnostic baselines. These results indicate that explicitly modeling feeder topology and equipment limits in data-driven frameworks is essential for reliable, dynamic electric vehicle hosting capacity assessment in adaptive distribution networks.

Semantics are one of the primary sources of top-down preattentive information. Modern deep object detectors excel at extracting such valuable semantic cues from complex visual scenes. However, the size of the visual input to be processed by these detectors can become a bottleneck, particularly in terms of time costs, affecting an artificial attention system's biological plausibility and real-time deployability. Inspired by classical exponential density roll-off topologies, we apply a new artificial foveation module to our novel attention prediction pipeline: the Semantic-based Bayesian Attention (SemBA) framework. We aim at reducing detection-related computational costs without compromising visual task accuracy, thereby making SemBA more biologically plausible. The proposed multi-scale pyramidal field-of-view retains maximum acuity at an innermost level, around a focal point, while gradually increasing distortion for outer levels to mimic peripheral uncertainty via downsampling. In this work we evaluate the performance of our novel Multi-Scale Fovea, incorporated into SemBA, on target-present visual search. We also compare it against other artificial foveal systems, and conduct ablation studies with different deep object detection models to assess the impact of the new topology in terms of computational costs. We experimentally demonstrate that including the new Multi-Scale Fovea module effectively reduces inherent processing costs while improving SemBA's scanpath prediction accuracy. Remarkably, we show that SemBA closely approximates human consistency while retaining the actual human fovea's proportions.

Current audio deepfake detection has achieved remarkable performance using diverse deep learning architectures such as ResNet, and has seen further improvements with the introduction of large models (LMs) like Wav2Vec. The success of large language models (LLMs) further demonstrates the benefits of scaling model parameters, but also highlights one bottleneck where performance gains are constrained by parameter counts. Simply stacking additional layers, as done in current LLMs, is computationally expensive and requires full retraining. Furthermore, existing low-rank adaptation methods are primarily applied to attention-based architectures, which limits their scope. Inspired by the neuronal plasticity observed in mammalian brains, we propose novel algorithms, dropin and further plasticity, that dynamically adjust the number of neurons in certain layers to flexibly modulate model parameters. We evaluate these algorithms on multiple architectures, including ResNet, Gated Recurrent Neural Networks, and Wav2Vec. Experimental results using the widely recognised ASVSpoof2019 LA, PA, and FakeorReal dataset demonstrate consistent improvements in computational efficiency with the dropin approach and a maximum of around 39% and 66% relative reduction in Equal Error Rate with the dropin and plasticity approach among these dataset, respectively. The code and supplementary material are available at Anonymous link.

Hyperparameter optimization (HPO) is computationally expensive due to large search spaces and dataset-dependent optimal configurations. To reduce this cost, transfer-based HPO methods have been proposed. However, existing methods rely solely on observed performance metrics, failing to disentangle the latent capability of hyperparameter configurations from dataset characteristics, which can cause biased transfer. In this paper, we propose a hyperparameter transfer method using Item Response Theory (IRT), a testing theory that enables the independent estimation of problem characteristics and examinee capability. By associating hyperparameter configurations with examinees and individual data samples with test items, the proposed method estimates the latent capability of configurations on source datasets. This is then combined with the dataset characteristics estimated from a small number of evaluations on a target dataset, enabling efficient hyperparameter transfer. Experimental results on multiple datasets demonstrate that the proposed method can select effective hyperparameter configurations with fewer evaluation trials compared to an existing method.

Representational similarity in neural networks is inherently scale-dependent, yet widely used metrics such as Centered Kernel Alignment (CKA) and Procrustes analysis provide only global scalar estimates. These scalars often fail to distinguish micro-scale geometric jitter (local noise) from macro-scale semantic reorganization, compressing multi-scale structural relationships into a single uninformative value. We introduce the Topological Alignment Spectrum (TAS), a multi-scale diagnostic tool that sweeps normalized mean Jaccard similarity over varying neighborhood sizes. By normalizing the metric over an analytically-derived expected range (from expected overlap under randomness to perfect alignment), TAS yields a dimension-invariant metric over a spectrum of scales, where one indicates perfect structural alignment, zero reflects chance-level agreement, and negative values signal active anti-alignment at specific scales. Experiments on synthetic point clouds demonstrate that TAS allows the recognition of distinct types of alignment perturbation: local jitter harms fine-grained neighborhoods but preserves cluster-level structure, while cluster-center shuffling preserves local similarity but disrupts global alignment -- phenomena that remain invisible or conflated under global, single-scalar metrics. Applying TAS to the MultiBERTs collection reveals that fine-tuning induces comprehensive topological reorganization across scales, challenging the view of task adaptation as merely conservative or localized. While models from different random seeds remain locally divergent, semantic clusters emerge as the dominant scale of alignment. TAS thus offers a granular, topology-aware alternative for diagnosing convergence and representational stability in deep networks.

Monte Carlo carrier-transport simulations used to estimate the spatial avalanche breakdown probability in single-photon avalanche diodes (SPADs) are essential for accurate photon detection probability (PDP) modeling, yet remain computationally prohibitive due to stochastic transport and rare impact-ionization events. A physics-informed cascade neural surrogate is introduced that emulates the Monte Carlo avalanche simulation workflow by decomposing it into three physically ordered stages: spatial electric-field reconstruction, excess-bias inference, and breakdown-probability prediction. The final stage is trained directly on Monte Carlo injection counts using a binomial negative log-likelihood, yielding calibrated spatial probability maps rather than mean-based regression outputs. Across variations in process, geometry, temperature, and bias, converged Monte Carlo breakdown-probability maps are accurately reproduced while runtime is reduced by by up to six orders of magnitude. More broadly, the cascade and likelihood-aligned training provide a reusable template for accelerating rare-event Monte Carlo in semiconductor transport and reliability, including high-voltage semiconductor device avalanche breakdown and radiation-induced degradation.

Multimodal Named Entity Recognition (MNER) and Multimodal Relation Extraction (MRE) extract entities and relations from text by fusing complex visual data. Despite notable advances, existing methods suffer from two core challenges: The first issue is the key information loss, where noise suppression eliminates entity-related visual cues in MNER and relation-critical details in MRE, degrading performance; The second issue is insufficient original information, which triggers entity misjudgment in MNER and relation determination deviations in MRE. This paper proposes a multimodal joint task network (AGLCLIE) to optimize MNER and MRE simultaneously. Specifically, an adaptive global-local contrastive learning module is designed to address the issue of key information loss: it constructs a Consistency-Aware Factor to quantify the semantic consistency between the textual and visual modalities and achieves adaptive and dynamic adjustment of temperature parameters in the two dimensions of "word-image patch" (global) and "phrase-region" (local). In addition, an information enhancement module is designed to address the problem of insufficient original information: it guides the Large Language Model to generate accurate explanations of entities and nouns through prompts, so as to supplement semantics and eliminate ambiguities. Experimental results show that our model achieves F1-scores of 75.97%, 87.82% and 87.27% on the Twitter15, Twitter17 and MNRE datasets, exceeding the prior state-of-the-art methods by 0.65%, 0.95% and 4.14%, respectively. These results confirm that AGLCLIE outperforms current SOTA methods on both MNER and MRE tasks.

Audio-visual navigation tasks require agents to locate and navigate toward continuously vocalizing targets using only visual observations and acoustic cues. However, existing methods mainly rely on simple feature concatenation or late fusion, and lack an explicit discrete representation of the target’s relative position, which limits learning efficiency and generalization. We propose Spatial-Aware Conditioned Fusion (SACF). SACF first discretizes the target’s relative direction and distance from audio-visual cues, predicts their distributions, and encodes them as a compact descriptor for policy conditioning and state modeling. Then, SACF uses audio embeddings and spatial descriptors to generate channel-wise scaling and bias to modulate visual features via conditional linear transformation, producing target-oriented fused representations. SACF improves navigation efficiency with lower computational overhead and generalizes well to unheard target sounds.

Audio-Visual Navigation (AVN) requires an em bodied agent to move toward a sounding target using both visual observations and binaural audio signals. However, in complex indoor environments, binaural cues are not always reliable. This problem becomes more noticeable when the agent encounters sound categories that were not seen during train ing. In this work, we present RAVN (Reliability-Aware Audio Visual Navigation), a framework that adjusts cross-modal fusion according to the reliability of audio observations. Instead of assuming that audio information is always trustworthy, our method estimates its reliability and uses this estimate to guide how audio and visual features are combined. To achieve this, we design an Acoustic Geometry Reasoner (AGR) trained with geometric proxy supervision. AGR learns observation-dependent uncertainty through a heteroscedastic Gaussian NLL objective. The learned uncertainty serves as a practical reliability signal and does not require geometric labels at inference time. Leveraging this signal, our Reliability-Aware Geometric Modulation (RAGM) module applies a soft gate to the incoming visual features. This explicitly suppresses inter-modal conflicts at the fusion stage. To test the architecture, we benchmarked RAVN within the SoundSpaces framework across Replica and Matterport3D layouts. Our comprehensive evaluations show consistent im provements over previous baselines, both in heard and unheard scenarios.

Audio-visual Navigation refers to an agent utilizing visual and auditory information in complex 3D environments to accomplish target localization and path planning, thereby achieving autonomous navigation. The core challenge of this task lies in the following: how the agent can break free from the dependence on training data and achieve autonomous navigation with good generalization performance when facing changes in environments and sound sources. To address this challenge, we propose an Audio Spatially-Guided Fusion for Audio-Visual Navigation method. First, we design an audio spatial feature encoder, which adaptively extracts target-related spatial state information through an audio intensity attention mechanism; based on this, we introduce an Audio Spatial State Guided Fusion (ASGF) to achieve dynamic alignment and adaptive fusion of multimodal features, effectively alleviating noise interference caused by perceptual uncertainty. Experimental results on the Replica and Matterport3D datasets indicate that our method is particularly effective on unheard tasks, demonstrating improved generalization under unknown sound source distributions.

Multimodal recommendation leverages item contents such as images and texts to alleviate interaction sparsity, yet it still faces noisy implicit feedback, unreliable content similarity, and entangled user intents. To address these issues, we propose BCMIRec, a behavior co-occurrence enhanced multi-interest dual-graph framework for multimodal recommendation. BCMIRec first constructs an item co-occurrence graph from co-selected behaviors and expands each user’s neighborhood to provide more behavior-consistent and clustered evidence. It then performs dual-graph propagation on the user--item bipartite graph and the co-occurrence graph, and fuses the resulting representations with an adaptive gate to balance global collaboration and local consistency. On top of the enhanced representations, BCMIRec learns multiple user intent vectors through temperature-controlled soft routing over the expanded neighborhood, and produces ranking scores via a LogSumExp-based mixture prediction with a diversity regularization to reduce interest collapse. Experiments on three Amazon benchmarks demonstrate that BCMIRec consistently outperforms strong multimodal baselines across Recall and NDCG, with clear gains on long-tail items.

Infrared and visible image fusion aims to integrate complementary information from multiple modalities. Recently, frequency-domain-based fusion methods have gained attention due to their advantages in edge texture recognition and global information integration. However, existing frequency-domain fusion methods fail to fully utilize the rich information embedded within frequency representations and the cross-modal complementarity among different spectral components. Therefore, we propose FHFusion. Specifically, our approach leverages the semantic and stylistic heterogeneity inherent in the frequency domain of infrared and visible images. Through spectral decoupling mechanisms, we achieve differentiated information fusion that effectively exploits cross-modal frequency domain complementary information. Further, we design an Amplitude Selection Module (ASM) and a Phase Discrimination Module (PDM) adaptively fusing energy distribution differences in the amplitude spectrum and structural complementary information in the phase spectrum, respectively. Experiments on M3FD, LLVIP and TNO dataset show FHFusion outperforms state-of-the-art methods in subjective and objective evaluations.

Accurate 3D object detection stands as a cornerstone of autonomous driving systems. Among various methodologies, the Bird's-Eye-View (BEV) paradigm, which fuses LiDAR point clouds and surround-view images, has emerged as the predominant approach. However, existing frameworks struggle to fully exploit the semantic potential of images and construct contextually integral LiDAR representations. To address these challenges, we propose FocusBEV, a "dual-focus" framework. First, we introduce the Semantic-Focus module, a 2D semantic guidance mechanism. Leveraging a lightweight 2D detector as a "semantic prior", this module proactively enhances foregrounds on feature maps prior to view transformation, thereby improving semantic fidelity. Second, we design the Long-range Interaction Aggregator (LIA) that uses shifted-window self-attention to reconstruct coherent global dependencies from sparse LiDAR features, overcoming the locality of CNNs. Finally, extensive experiments conducted on nuScenes demonstrate that FocusBEV achieves 72.3 mAP and 74.6 NDS, explicitly validating the effectiveness of our FocusBEV.

Video Temporal Grounding (VTG), which aims to localize video clips corresponding to natural language queries, is a fundamental yet challenging task in video understanding. Existing Transformer-based methods often suffer from redundant attention and suboptimal multi-modal alignment. To address these limitations, we propose MLVTG, a novel framework comprising two key designed modules: MambaAligner and LLMRefiner. MambaAligner uses the bidirectional scanning and gated filtering strategy to model temporal dependencies and extract robust video representations for multi-modal alignment. LLMRefiner leverages the specific frozen layer of a pre-trained Large Language Model (LLM) to implicitly transfer semantic priors, enhancing multi-modal alignment without fine-tuning. This dual alignment strategy, temporal modeling via structured state-space dynamics and semantic purification via textual priors, enables more precise localization. Extensive experiments on QVHighlights, Charades-STA, and TVSum demonstrate that MLVTG achieves highly competitive performance.

The accurate diagnosis of liver lesions from 3D multimodal Magnetic Resonance Imaging (MRI) is a critical yet challenging task in oncological decision-making. Existing deep learning approaches often struggle due to the scarcity of annotated volumetric medical data and the inherent heterogeneity of MRI sequences, where distinct physical parameters (e.g., T2-weighted vs. DWI) create significant distributional shifts. Furthermore, current medical Vision-Language Models (VLMs) adapted for 3D tasks often rely on \textbf{static early fusion strategies}, treating diverse MRI modalities as homogeneous input channels. This indiscriminate integration leads to feature interference and fails to replicate the \textbf{dynamic, semantic-driven} nature of radiological interpretation. To address these limitations, we propose \textbf{MIRAGE} (\textbf{M}ultimodal MRI \textbf{R}epresentation \textbf{A}ggregation with \textbf{G}ated \textbf{E}xperts). We introduce a library of modality-aware expert encoders, adopting a \textbf{decoupled learning paradigm}: experts are first contrastively pretrained to align specific physical imaging features with LLM-synthesized radiology semantics, ensuring robust representation learning prior to diagnosis. To mimic the selective attention of radiologists, we design a \textbf{Modality Routing Gate (MRG)} that dynamically prioritizes the most discriminative expert path for each patient, reducing redundancy and gradient conflicts. Additionally, a Cross-modal Interactive Feature Extraction and Refinement (CIFER) module is employed to fuse these selective features with textual knowledge. Extensive evaluations on the large-scale LLD-MMRI dataset demonstrate that MIRAGE achieves state-of-the-art performance with 91.57\% accuracy, significantly outperforming existing medical VLMs while maintaining computational efficiency suitable for clinical deployment.

While modern multi-modal models rely on dense visual-language alignment, the extent to which pure text-only Large Language Models (LLMs) can perform visual reasoning remains underexplored. We investigate this by presenting DeepYOLO, a neuro-symbolic framework that converts images into compact, grammar-constrained symbolic scene reports using an off-the-shelf object detector. This allows a text-only LLM (e.g., DeepSeek-V3) to answer visual questions without any multi-modal pre-training. Our experiments on a VQAv2 validation subset show that DeepYOLO achieves 42.3% accuracy, significantly outperforming an explicit Scene Graph baseline (38.2%) while requiring 2.5× fewer tokens. Crucially, replacing detector outputs with ground-truth objects (Oracle) yields 44.5%, demonstrating that the symbolic interface is sufficient for the task and that the primary bottleneck remains the object detector rather than the LLM’s reasoning capability. DeepYOLO offers a modular, auditable, and highly token-efficient alternative for visual reasoning, particularly suitable for privacy-sensitive or edge-compute scenarios.

Weakly supervised video anomaly detection (WSVAD) aims to identify abnormal events in long, untrimmed videos using only video-level labels, which is both practical and challenging. Existing methods often suffer from semantic ambiguity, caused by the diverse and context-dependent nature of anomalies, and attention dispersion, where models fail to focus on subtle anomalous cues. To address these issues, we propose VAPrompt, a semantic distribution-aware framework that explicitly models anomaly uncertainty and enhances spatiotemporal localization. The core component, Variational Text Prompt Refinement (VTPR), represents textual prompts as stochastic latent distributions rather than fixed embeddings, enabling the model to capture diverse anomaly semantics and improve generalization. In addition, a Prompt-Infused Spatiotemporal Saliency Aggregation (PISSA) module integrates cross-modal attention to highlight anomaly-relevant regions while suppressing background noise. Furthermore, a Multi-Anchor Semantic Alignment Objective (MASAO) is introduced to enforce fine-grained semantic separation under weak supervision. Extensive experiments on the UCF-Crime and XD-Violence benchmarks demonstrate that VAPrompt consistently outperforms state-of-the-art methods. Specifically, our approach achieves 89.01% AUC and 72.32% AnoAUC on UCF-Crime, and 86.16% AP on XD-Violence, along with substantial improvements in temporal localization accuracy. These results indicate that modeling semantic uncertainty and integrating prompt-guided saliency are effective for robust weakly supervised video anomaly detection, making VAPrompt a practical and extensible solution for complex real-world surveillance scenarios.

From the creation of early mobile network standards to current 5G and beyond, interference management has been extensively researched as it is crucial for improving network performance and service quality. Among the different techniques, downlink power control, which consists of dynamically adjusting the power with which base stations transmit to users, has demonstrated to be critical to mitigate inter-cell interference. In this context, through the years, traditional optimization methods have been applied, achieving good performance but facing challenges in current standards characterized by high complexity in terms of service requirements. In this paper, we propose a deep reinforcement learning based approach for downlink power control, motivated by the results achieved in recent years by data-driven approaches to radio resource management optimization. We train an agent based on the double dueling deep Q-network algorithm, relying on a centralized single-agent reinforcement learning architecture. The OMNeT++ discrete-event simulator with the Simu5G library is chosen to simulate a 5G Standalone scenario where the agent is trained with the goal of maximizing the network-wide signal-to-interference-plus-noise ratio. We tested the trained agent in this scenario against the baseline approach native to the simulator, based on fixed power transmission. Validation results show that the proposed approach improves the signal-to-interference-plus noise ratio while achieving a reduction in physical resource block usage compared with the baseline approach. While the work presented in this paper is based on scalar network key performance indicators, it establishes a baseline for future improvements to incorporate multimodal data, enabling the use of multimodal architectures to achieve a deeper understanding of the network state.

Elevated PM10 concentrations, driven by both anthropogenic emissions and frequent sand and dust events, pose serious threats to public health and infrastructure; accurate PM10 forecasting is therefore essential for pollution control and resilient energy operations. This study proposes Dustformer, a spatiotemporal forecasting framework designed to fuse ground observations, meteorological data, satellite remote sensing data, and topography data. The proposed architecture integrates a Mixture-of-Experts (MoE) Transformer with dynamic routing to capture complex multimodal temporal patterns, coupled with an attention-weighted Dynamic Graph Neural Network (DGNN) to explicitly model evolving inter-station pollutant diffusion paths. Additionally, a robust sample screening mechanism is incorporated to mitigate the impact of data noise. Extensive experiments on real-world datasets demonstrate that Dustformer outperforms several representative baseline methods in terms of predictive accuracy and robustness. Furthermore, systematic ablation studies confirm the indispensability of each module, revealing that adaptive temporal expert routing and dynamic spatial dependency learning are critical for minimizing forecasting errors.

Real-world image degradations pose severe challenges for infrared-visible image fusion (IVIF), as most existing methods rely on the implicit assumption of pristine inputs. Conventional single-teacher frameworks struggle to balance image restoration and detail preservation, as denoising objectives inevitably suppress the high-frequency textures critical for fusion. Similarly, cascaded two-stage approaches suffer from error propagation and limited end-to-end efficacy. To resolve this fundamental conflict, we propose a novel Decoupled Dual-Teacher (D2T) framework that disentangles restoration and fusion into dedicated expert networks. A lightweight student network distills this knowledge via a Quadruple-Supervision Strategy, aligning its intermediate features with robust representations from both teachers to achieve simultaneous restoration and fusion. Notably, our model operates in a fully blind manner, handling unknown degradations without requiring manual priors while maintaining high inference efficiency. Extensive experiments on challenging degraded datasets show that D2T consistently outperforms state-of-the-art (SOTA) fusion methods, producing fused images with superior clarity and structural fidelity while effectively suppressing noise and artifacts from unknown degradations.

Large language models (LLMs) are increasingly used to support language-based interaction in autonomous driving, including scene question answering and risk explanation. Yet a critical reliability failure persists: when prompted with long, redundant, and mutually inconsistent perception logs from heterogeneous sensors, LLMs may produce fluent but unsupported statements, hallucinate non-existent traffic agents, and amplify false premises. We address this problem from an interface perspective. We present CoherentDrive, a prompt-ready grounding framework that compiles multi-source perception facts into a compact, conflict-resolved Coherent World State (CWS) with entity identifiers, provenance, and uncertainty flags. CoherentDrive further enforces a grounding-first answering protocol that restricts responses to CWS-supported entities and relations, enabling explicit refusal when evidence is missing. On nuScenes-QA under a zero-shot setting with multiple LLM backbones, CoherentDrive consistently improves answer correctness while substantially reducing hallucinated entity mentions, and in a single-pass CWS prompting setting it also shortens prompt length and reduces inference latency. Extensive experiments demonstrate that conflict-resolved world-state prompting provides an effective and reliable foundation for hallucination-resistant driving question answering.

Improving clustering performance on incomplete multi-view data is crucial for intelligent systems. However, most existing incomplete multi-view clustering methods assume consistent missing patterns across views, which rarely holds in real-world applications with heterogeneous and view-specific missingness. To address this issue, we propose SIGMA (Self-supervised Graph Contrastive Learning for Incomplete Multi-view Data Alignment), a method that enables effective multi-view representations under inconsistent missing patterns. Compared with existing studies, SIGMA extracts view-specific structures while independently inferring missing information for each view. Specifically, SIGMA initializes view-specific graphs and a global consistent graph using $k$-nearest neighbors, and dynamically propagates reliable structural information from the global graph to incomplete views. Furthermore, a weighted graph contrastive learning strategy is introduced to preserve both local discriminability and global structural consistency across recovered views. Extensive experiments on multiple benchmark datasets demonstrate that SIGMA consistently outperforms state-of-the-art incomplete multi-view clustering methods.

Cross-modal matching relies on large-scale data, which researchers often collect through crowd-sourcing or web crawling to reduce annotation costs. However, this process inevitably introduces mismatched pairs, leading to the noisy correspondence problem. Since noisy correspondences are difficult to identify and correct, noise-robust cross-modal learning becomes a challenging task. To achieve robust cross-modal retrieval, we propose a noisy label re-matching and pseudo-classification framework. The noisy label re-matching module identifies and corrects mismatched pairs by estimating correspondence confidence in the learned representation space and refining alignment with a pre-trained vision–language model. The pseudo-classification module treats titles as category labels and generates pseudo-labels from clean samples to promote confident and balanced predictions. In addition, a confidence-based weighting strategy is introduced to adaptively suppress harmful samples during training. Extensive experiments on Flickr30K, MS-COCO, and CC120K demonstrate that our method outperforms state-of-the-art approaches.

Accurate matching between SAR and optical images remains challenging due to the significant discrepancy in imaging mechanisms and visual characteristics.To address this issue, this paper proposes a cross-modality knowledge transferring framework for SAR--optical image matching.First, an adversarially accelerated diffusion model is introduced to construct an intermediate domain between SAR and optical images, effectively bridging the modality gap.Based on the translated representations, a cascaded cross-sensor matching architecture is developed using a densely connected convolutional network equipped with hierarchical attention mechanisms, including cross-channel, cross-tensor, and cross-kernel attention.Extensive experiments on measured datasets demonstrate that the proposed method significantly outperforms competitive approaches.

Efficient segmentation of medical images has evolved from relying solely on Convolutional Neural Networks (CNNs) to exploring hybrid models that integrate CNNs with Vision Transformers (ViTs). Despite the advantages of transformers in capturing global dependencies, their high computational and storage demands pose challenges. This research introduces the novel ViL-UNet, which combines CNNs with Vision Extended Long Short-Term Memory (Vision-xLSTM), aiming to balance performance and computational efficiency. The Vision-xLSTM blocks capture long-range spatial and global relationships, both at the network’s bottleneck and within the skip connections to refine multi-scale feature fusion. Our primary goal is to demonstrate that Vision-xLSTM provides a suitable backbone for medical image segmentation, offering superior performance with reduced computational costs. The ViL-UNet achieves state-of-the-art results on the publicly available Synapse and ACDC datasets. Index Terms— Medical Image Segmentation, xLSTM, 3D Imaging.

Pleural effusion, as a highly prevalent clinical symptom, has garnered wide-spread attention. However, most models are designed for lung cancer detection, but the detection of tumor lesions within pleural effusion still requires improvement. This paper proposes a lung cancer detection model integrating orthogonal views and blood analysis indicators for patients with pleural effusion(OVBI-LCDM), in which feature enhancement network based on view complementarity is designed to more accurately depict hidden lesions in orthographic views. Blood indicators analysis module is introduced to analyze correlation and importance of six blood indicators and determine their weights. Feature fusion module based on analysis results is proposed to dynamically re-evaluate the contributions of orthogonal view images and blood indicators, thereby achieving mutual enhancement of these features. The experimental results show that the model achieved good classification performance, with an accuracy of 82.76%, outperforming other models.

Chinese cultural heritage artifacts are rich in multimodal signals, making image-text cross-modal retrieval crucial for digital preservation and reuse. Yet reliable image-text alignment remains challenging due to (i) background-heavy professional descriptions that introduce redundant semantics and (ii) CLS-level global matching that can obscure locally discriminative visual evidence. We propose EGF-CHITR, a lightweight framework that integrates entity-prior gating with bidirectional local alignment to enhance cross-modal discriminability. Specifically, entity-prior gating injects domain knowledge to reweight textual tokens, while bidirectional local alignment models hierarchical patch--token interactions to suppress redundant semantics and strengthen locally discriminative correspondence under global semantic consistency. Experiments on the ceramics dataset PorTi and the public cultural-heritage dataset CulTi demonstrate that EGF-CHITR consistently outperforms mainstream methods across multiple metrics in both retrieval directions.

Audio-visual multimodal person verification has attracted increasing attention for improving robustness in challenging conditions by leveraging the complementarity between speech and face cues. Despite recent progress, existing feature-level fusion methods---e.g., soft-attention and gating---typically infer modality reliability implicitly from identity embeddings, without dedicated degradation-aware cues for acoustic noise or facial occlusion. Consequently, the learned modality weights can be unreliable when degradations are severe or asymmetric, yielding suboptimal fusion representations. In this work, we propose a mask-guided adaptive fusion framework (MGAF) that explicitly estimates modality reliability from degradation-aware masks and uses this signal to guide fusion. Specifically, MGAF predicts a time--frequency mask to characterize noise-corrupted speech and learns a feature-level visual mask on convolutional representations to enhance reliable responses; the masks are aggregated into explicit reliability cues, based on which MGAF performs mask-driven re-weighting of unimodal embeddings and constructs the fused representation via weighted concatenation. This design allows the model to emphasize the more reliable modality under strong corruption while retaining complementary information when both modalities are informative. We train MGAF on VoxCeleb2 and evaluate it on VoxCeleb1 under systematically simulated noise and occlusion settings, including asymmetric degradations. On the standard VoxCeleb1 test sets, MGAF achieves EERs of 0.207\%, 0.212\%, and 0.433\% on VoxCeleb1-O, VoxCeleb1-E, and VoxCeleb1-H, respectively, and achieves consistent gains over representative fusion baselines under degraded scenarios---especially under asymmetric degradations---without over-suppressing either modality when both are reliable.

Multi-modal object Re-Identification (ReID) exploits complementary cues from heterogeneous sensors to recognize identities under challenging environmental conditions. However, existing methods suffer from three key limitations. First, modality reliability can vary substantially across instances and environments, rendering static fusion suboptimal. Second, sensor-induced spatial misalignment often corrupts cross-modal correspondences and degrades feature fusion. Third, naive feature aggregation lacks a progressive mechanism to integrate identity cues from fine-grained details to high-level semantics. To address these issues, we propose AdapDeFormer, a unified framework that combines Enhanced Deformable Attention (EDA) with a Hierarchical Adaptive Query Network (HAQN) for robust multi-modal representation learning. Specifically, EDA reweights modalities via Adaptive Modal Weighting (AMW) and corrects spatial discrepancies through Multi-Scale Offset Fusion (MSOF) to enable alignment-aware fusion. Building on the enhanced features, HAQN performs progressive identity feature extraction with hierarchical query learning, where cross-layer query propagation and modality-adaptive aggregation facilitate discriminative representation learning. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness and robustness of AdapDeFormer.

Chart Question Answering (ChartQA) has recently gained significant attention in visual data understanding. Despite achieving promising results on standard benchmarks, existing vision-language models—ranging from lightweight architectures to recent Large Multimodal Models (LMMs)—often degrade significantly under diverse chart styles and visual disturbances, such as variations in color, font, and layout. To improve robustness, we propose ChartMVC(Chart Multi-View Consistency Contrastive learning), a general Multi-View Contrastive Learning framework designed to enhance visual stability in chart understanding. ChartMVC forces the model to learn semantically consistent representations across diverse visual styles through a queuebased contrastive mechanism. To validate the effectiveness of our framework, we instantiate it using the UniChart architecture. Extensive experiments on the ChartQA benchmark show that our approach significantly outperforms the original baseline in relaxed accuracy metrics. Furthermore, the model exhibits superior robustness against various chart perturbations, demonstrating the effectiveness of the ChartMVC framework for real-world chart understanding tasks.

As a core component of autonomous driving perception systems, 3D object detection faces a critical bottleneck: how to achieve effective joint representation of geometric features and semantic information across heterogeneous modalities (LiDAR point clouds and RGB images). Prevailing multi-sensor fusion methods are prone to cascading errors stemming from sequential optimization processes spanning geometric modeling, cross-modal alignment, and semantic integration. These errors ultimately lead to systematic failures in semantic understanding and undermine robustness. To address this limitation, we propose the UniVerse framework, which systematically cascades a progressive optimization paradigm through "feature disentanglement, alignment enhancement, and deep aggregation." Specifically, UniVerse introduces a novel triple synergy mechanism: the unified dual-stream perception encoder (GeSe-Encoder) ensures high-purity geometric and semantic features at the source; the cross-modal hierarchical feature fusion module (CHFM) achieves precise alignment through bidirectional enhancement; and the semantic-aware global enhancement module (SAGE) enables robust semantic understanding through multi-source interaction, ultimately supporting effective joint modeling of geometric structure and semantic context. Extensive experiments on the KITTI dataset validate the effectiveness of our method, achieving 65.63% 3D mAP and outperforming comparable baselines by 3.52%, confirming the value of the proposed cascade optimization strategy.

Social media platforms have become a primary source for mining public sentiment and analyzing user behavior. While explicit emotions are easily detectable, a significant portion of online expression is implicit—conveyed through subtle metaphors, irony, or context—posing a severe challenge for traditional sentiment analysis models. Existing approaches typically rely on surface-level semantic features, often failing to capture these underlying emotional cues due to a lack of background knowledge and reasoning capabilities. To address this, we propose a novel framework: a Social Media Emotion Detection Model based on Implicit Emotion Multi-Hop Memory and Knowledge Embedding. Unlike standard methods, our model incorporates a Knowledge-Embedded Attention Mechanism to retrieve external commonsense knowledge, bridging the gap between literal text and implicit intent. Furthermore, we design an Implicit Emotion Multi-Hop Memory Network to capture long-range contextual dependencies and learn the prior distribution of emotional states. Experiments on the fine-grained GoEmotions dataset demonstrate that our proposed method achieves state-of-the-art performance, significantly outperforming strong baselines in detecting implicit and complex emotion categories.Experiments on the fine-grained GoEmotions dataset demonstrate that our proposed method achieves state-of-the-art performance, yielding a Precision of 54.66%, Recall of 53.80%, and Macro-F1 score of 52.34%. These results significantly outperform strong baselines such as BERT (Macro-F1: 46%) and BiDLSTM (Macro-F1: 50.4%). This superior performance is attributed to our model's unique ability to bridge implicit semantic gaps via external knowledge embeddings and capture long-range emotional dependencies through the multi-hop memory network.

Unaligned multimodal medical image fusion faces dual challenges of spatial misalignment and cross-modal information inconsistency. Existing collaborative registration–fusion frameworks often enforce global alignment, causing unreliable regions such as low-texture areas and noise to severely interfere with deformation estimation; they are also prone to inducing erroneous deformations and ghosting in regions with large modality discrepancies. To address these challenges, from the perspective of distinctly handling anatomical misalignment that should be aligned and modality-specific differences that should be preserved, we propose EDARFusion: Entropy-Difference-Aware Dynamic Registration for Unaligned Multimodal Medical Image Fusion. Specifically, the Feature-Reliability Guided Pre-Registration Adapter (FRG-PreReg) predicts a feature-level reliability prior map to suppress the interference of unreliable regions in deformation learning; the Entropy-Difference-Aware Flow Registration (EDA-FlowReg) generates a gating-weight map from local information entropy differences and imposes soft constraints on incremental deformation updates in multi-scale estimation, thereby enabling discrepancy-aware dynamic alignment and reducing ghosting and fusion artifacts caused by misregistration; in the fusion stage, an alignment-consistency constraint is introduced to dynamically integrate structural and modality-specific information on the basis of reliability-aware alignment. Extensive experiments and visualizations of the local information entropy-difference map, the gating-weight map, and the reliability map demonstrate that the proposed method achieves superior registration and fusion performance compared with multiple representative approaches, with interpretable decision-making.

With the rapid development of embodied intelligence, robust image classification from low-resolution multimodal sensory inputs has become a critical objective for embodied agents. Due to hardware limitations and environmental interference, image inputs captured by embodied agents often suffer from severe resolution degradation, and manually designed architectures for low-resolution classification lack adaptability to embodied multimodal perception. To address these challenges, we propose an \emph{Embodied Multimodal Neural Architecture Search} (EM-NAS) system, which automates the design of network architectures for low-resolution image classification in embodied scenarios. Specifically, we design an embodied perception-oriented NAS search space that integrates super-resolution (SR) as a feature enhancement component and incorporates cross-modal fusion operators, enabling the automatic discovery of architectures tailored to embodied classification demands. We further propose a Context-Aware Multimodal Fusion (ECMCF) module, which is optimized via NAS to dynamically fuse visual and environmental modal features for robust feature representation. Additionally, an embodied joint loss function is introduced to guide NAS search toward architectures with both high classification accuracy and environmental robustness. Experiments on extended Embodied-CIFAR10 and Tiny-ImageNet datasets demonstrate that the EM-NAS-discovered architecture outperforms state-of-the-art baselines, achieving a 4.2\% improvement in classification accuracy (with auxiliary SR gains of 2.5 dB in PSNR and 0.027 in SSIM). This work introduces a NAS-driven paradigm for embodied low-resolution image classification, providing an automated architecture design solution for embodied perception systems.

The accurate prediction of three-dimensional power distribution is critical for ensuring the safety of nuclear reactors. Traditional prediction methods struggle in computational efficiency and prediction accuracy. Recent deep learning approaches offer a promising alternative but frequently employ generic architectures due to a lack of domain-specific knowledge guiding the architecture design. They fail to capture the inherent anisotropy, geometric symmetry, and state-dependent nonlinearity of reactor behavior. To overcome each limitations, this paper proposes a conditional neural architecture search method. It comprises three essential components: (1) an anisotropic convolution block is designed to decouple radial and axial feature learning, thereby matching the anisotropy of physical principles; (2) a geometry-aware attention block is proposed to embed rotational symmetry priors and adaptively focus on critical regions; and (3) a lightweight conditioner is developed to dynamically activate the architecture for specific reactor states, effectively modeling state-dependent nonlinearities. Experiments show that our automatically searched architecture significantly outperforms manual baselines. For instance, it achieves a mean square error of 0.0109 with only 873K parameters, reducing error by 44.1% compared to a V-Net baseline. The results demonstrate that our method enables highly accurate, efficient, and adaptive prediction suitable for reactor safety evaluation.

Standard graph neural networks (GNNs) typically perform message passing over all nodes during training, incurring substantial computational redundancy. To solve the problem, this paper proposes SmartDyGNN, a dynamic sparse training algorithm that adaptively activates critical nodes by jointly exploiting structural redundancy in graph-structured data, embedding dynamics, and task-gradient sensitivity. Embedding dynamics capture the convergence trends of node representations throughout training, while task-gradient sensitivity reflects how much each node’s representation influences the final task loss. Together, these signals guide the algorithm in dynamically identifying and retaining the most optimization-relevant nodes at each training iteration. On benchmark datasets Cora, Citeseer, and PubMed, the proposed method stably maintains the proportion of active nodes at approximately 73%,thereby reducing message-passing operations by 25%–30%, while achieving a relative improvement of 0.20%–0.37% in accuracy over the full-graph baseline. Moreover, SmartDyGNN generalizes well to other architectures, including GCN and GAT.

Neural Architecture Search (NAS) has become a pivotal technique in automated machine learning. Evolutionary Algorithm (EA)-based methods demonstrate superior search quality but suffer from prohibitive computational costs, while gradient-based approaches like DARTS offer high efficiency but are prone to premature convergence and performance collapse. To bridge this gap, we propose G-ICSO-NAS, a hybrid framework implementing a three-stage optimization strategy. The Warm-up Phase pre-trains supernet weights w via differentiable methods while architecture parameters α remain frozen. The Exploration Phase adopts a hybrid co-optimization mechanism: an Improved Competitive Swarm Optimizer (ICSO) with diversity-aware fitness navigates the architecture space to update α, while gradient descent concurrently updates w. The Stability Phase employs fine-grained gradient-based search with early stopping to converge to the optimal architecture. By synergizing ICSO's global navigation capability with differentiable methods' efficiency, G-ICSO-NAS achieves remarkable performance with minimal cost. In the context of the DARTS search space, an accuracy of 97.46% is achieved on CIFAR-10 with a computational budget of just 0.15 GPU-Days. The method also exhibits strong transfer potential, recording accuracies of 83.1% (CIFAR-100) and 75.02% (ImageNet). Furthermore, regarding the NAS-Bench-201 benchmark, G-ICSO-NAS is shown to deliver state-of-the-art results across all evaluated datasets.

In modern manufacturing systems, dual-resource scheduling has emerged as a critical challenge, where each operation must be simultaneously assigned to both an eligible machine and a qualified worker. This joint allocation requirement introduces tight resource coupling, which significantly expands the combinatorial search space while severely constraining the feasible solution region and resulting in complex multi-dimensional trade-offs between completion time and operational costs. This paper investigates a Dual-Resource Constrained Flexible Job Shop Scheduling Problem (DRC-FJSP) that minimizes the makespan and a comprehensive production cost integrating raw-material holding, work-in-process holding, direct processing costs, and earliness/tardiness penalties. To effectively approximate Pareto-optimal solutions, we propose EM-MFF-MOPSO, a discrete multi-objective framework. The algorithm features an entropy-guided mechanism for stage-based parameter adaptation, a Maximin Fitness Function (MFF) strategy for leader selection, and a Simulated Annealing (SA) local search to refine solutions. Experiments on reproducible benchmark instances of increasing scales demonstrate that EM-MFF-MOPSO consistently produces more competitive and better-distributed Pareto fronts than representative baselines.

Data-driven evolutionary algorithms (DDEAs) have been widely used for solving expensive optimization problems by replacing expensive fitness evaluations with surrogate models constructed from limited evaluated data. However, although various surrogate construction and selection strategies have been investigated, the evaluation scale of surrogate-assisted fitness estimation is usually predefined and remains unchanged throughout the evolutionary process. Such a static evaluation scale may be inefficient or even misleading, as the evaluation requirements vary across different evolutionary stages. To alleviate this issue, this paper proposes a scale-adaptive model management (SAMM)-based DDEA, named SAMM-DDEA, together with three contributions. First, a SAMM strategy is proposed to dynamically adjust the evaluation scale of surrogate-assisted fitness estimation according to evolutionary feedback. In addition, a Full-Coverage Sampling (FCS) strategy is proposed to enhance ensemble surrogate diversity under limited data conditions, so as to enhance algorithm performance. Third, the SAMM-DDEA is developed and validated on widely-used expensive optimization problems with 10 to 100 dimensions, demonstrating superior or highly competitive performance compared with several state-of-the-art DDEAs.

In recent years, numerous surrogate-assisted evolutionary algorithms (SAEAs) have been developed for expensive multi-objective optimization problems (EMOPs). However, under a stringent evaluation budget, many existing SAEAs are hindered by low-quality initialization, inefficient use of training data for surrogate modeling, and limited exploitation of historical high-quality solutions. This paper proposes an SAEA with geometry-guided initialization and memory-guided variation (GIMV-SAEA). Specifically, a geometry-guided initialization scheme combines boundary probing with a center anchor and early local elite sampling to generate high-quality initial solutions while maintaining uniform exploration. An evolutionary-stage-aware adaptive training set construction strategy is further designed to balance surrogate training cost and predictive accuracy. In addition, a memory-guided variation operator preserves historical elite solutions and applies stronger perturbations to poorly converged decision variables. Experimental results on the DTLZ, WFG, and MaF test suites, as well as offset DTLZ instances, demonstrate that GIMV-SAEA achieves competitive performance on expensive multi- and many-objective optimization.

Most existing algorithms for large-scale multiobjective optimization problems (LSMOPs) suffer from monotonous and fixed search strategies and evolutionary stages. Consequently, the solutions generated by these algorithms struggle to achieve rapid convergence toward the Pareto front. To address this issue, this paper proposes a Multi-stage Large-scale Multiobjective Evolutionary Algorithm integrating Multi-directional sampling and Branch Search (LMOEA-MBMS). Specifically, a breadth multi-directional sampling strategy is first introduced to explore five distinct directions correlated with the sample solutions. Subsequently, a depth branch search strategy is implemented to further explore the non-dominated solutions of these samples along the sampling directions, thereby enhancing the algorithm’s convergence. Finally, a multi-stage environmental selection framework is designed to achieve the specific goals of the algorithm across different evolutionary phases. The proposed algorithm was evaluated against six state-of-the-art algorithms on eighteen LSMOPs and two practical problems of Time varying Ratio Error Estimation (TREE). Experimental results demonstrate that the performance of the proposed algorithm outperforms other state-of-the-art methods, effectively improving both convergence and solution diversity, with significant advancements in both theory and application.

Neural architecture search (NAS) aims to automatically discover high-performing network architectures for specific tasks. In recent years, NAS methods that leverage sharpness signals have shown promising potential for improving model generalization. However, NAS still falls short in robustness under out-of-distribution (OOD) shifts. Recent studies have incorporated Sharpness-Aware Minimization (SAM) into the search process, but under weight-sharing and short training-budget settings, the standard SAM perturbation direction may entangle deterministic gradient trends with mini-batch stochastic noise, thereby biasing the sharpness proxy. To address this issue, we propose GFNAS, a gradient-friendly sharpness-aware NAS framework that estimates the deterministic gradient trend via an exponential moving average (EMA) and suppresses it when constructing perturbations, yielding a noise-consistent direction. Without increasing the search cost or modifying the search space, this strategy provides a more robust robustness-oriented evaluation signal. Experiments on CIFAR-10/100 and their corrupted variants demonstrate that GFNAS improves both accuracy and corruption robustness (mCE), reduces search cost, and alleviates sensitivity to sharpness hyperparameters.

Traffic flow prediction has been recognized as a pivotal technology in the development of smart cities, facilitating data support for resource scheduling and traffic management. By leveraging an efficient spatiotemporal graph attention mechanism, outstanding performance in balancing prediction accuracy and computational efficiency has been demonstrated by the STGformer model. However, a static adjacency matrix fails to accommodate dynamic changes in road connectivity within actual traffic scenarios, such as road closures or peak-hour congestion, consequently limiting prediction performance under complex conditions. To address these limitations, an enhanced STGformer model, termed ADG-Former, is introduced based on adaptive optimization of dynamic graph structures. Initially, a dynamic adjacency matrix is constructed using spatiotemporal correlations of traffic flow and speed, wherein effective connectivity relationships between nodes are updated in real time through an adaptive threshold strategy. Subsequently, a gated fusion module is employed to dynamically balance the contribution weights of static road topology and dynamic traffic dependencies, enhancing adaptability to complex scenarios. This study presents a feasible optimization scheme for traffic flow prediction in dynamic scenarios and demonstrates engineering applicability.

The deployment of Large Machine Learning Models (LMLMs) on edge devices is hindered by the conflict between the colossal size of modern architectures and the limited, heterogeneous resources of edge clients. Federated Learning (FL) enables privacy-preserving collaborative training, yet standard FL protocols typically assume model homogeneity, making them unsuitable for diverse edge environments. In this paper, we propose Auto-HFL, a framework for Automated Sub-Network Discovery. Unlike traditional Neural Architecture Search (NAS) which incurs high computational costs, Auto-HFL introduces a lightweight, resource-aware discovery mechanism. It allows clients to automatically extract and train optimal sub-networks from a global supernet based on their real-time hardware constraints (e.g., memory and bandwidth). We further design a specialized aggregation strategy to handle the structural heterogeneity of the discovered models. Extensive experiments on [Insert Dataset, e.g., CIFAR-10/100] demonstrate that Auto-HFL significantly reduces communication and computation overhead while maintaining accuracy comparable to full-model training. This work paves the way for automating model deployment in the era of large-scale edge intelligence.

Evolutionary multi-task optimization (EMTO) has been widely applied to address the challenges of high-dimensional feature selection. However, existing EMTO algorithms still suffer from two key deficiencies: in the feature search process, the flipping strategy uses fixed probabilities and thresholds, failing to adapt to population dynamics during evolution; and in the knowledge transfer stage, the adaptive aggregation strategy randomly selects sources, ignoring particle value and task correlation, which may lead to negative transfer. To overcome these limitations, this paper proposes AAMCSO-AP, a dynamic adaptive aggregative multi-task competitive particle swarm optimization algorithm. It introduces two novel mechanisms: a population entropy and feature distribution-aware dynamic update strategy for winners, and a value assessment and relationship-aware knowledge transfer strategy for losers, enabling adaptive parameter adjustment and intelligent knowledge transfer. Experiments on seven real high-dimensional datasets (with up to 10,509 features) demonstrate that AAMCSO-AP achieves significantly superior classification accuracy and feature reduction compared to the original AAMCSO and other state-of-the-art algorithms, confirming the effectiveness of the proposed dynamic mechanisms.

Graph Neural Networks (GNNs) have achieved remarkable success in processing graph-structured data, yet they continue to face two major limitations: high computational cost and the presence of structural noise in real-world graphs, such as missing edges or spurious connections. Existing sparse training approaches effectively reduce computational overhead by pruning model weights and node features. However, they operate strictly on the initial graph topology, and the resulting trained graph is merely a subgraph of the original one, leaving the structural optimization problem unaddressed. By contrast, Graph Structure Learning (GSL) methods are capable of dynamically optimizing graph topology during training. However, these approaches typically rely on global pairwise similarity computation, resulting in a quadratic time complexity of O(N^2), which severely limits their scalability to large-scale graphs. Although anchor-based variants of Iterative Deep Graph Learning (IDGL) [1] partially alleviate this issue, they depend on low-rank approximations that inevitably discard local structural information of the original graph and may further introduce additional optimization instability. Therefore, an efficient graph structure optimization method that preserves local structural details is highly desirable. To address these limitations, we propose a dual-sparsity framework that integrates the principles of IDGL with the Comprehensive Graph Gradual Pruning (CGP) [2] paradigm, enabling the joint optimization of model parameters and graph topology. Specifically, we compute cosine similarity between node pairs to measure semantic relevance, which guides both the addition of necessary edges and the removal of redundant ones. Meanwhile, to overcome the high computational cost inherent in standard IDGL, we employ a hybrid sampling strategy that combines 2-hop neighborhood search with random negative sampling, reducing the computational complexity from O(N^2) to O(E). This strategy also mitigates the risk of the optimization process falling into local minima. Experiments on the Cora, Citeseer, and Pubmed datasets demonstrate that the designed framework achieves good performance while maintaining dual sparsity in both the model and graph structure, achieving comparable or superior performance static sparse baselines and exhibiting particularly notable generalization improvements on the larger Pubmed dataset.

Deep learning-based encrypted traffic classification often relies on fixed network architectures with simple fusion strategies for multi-modal traffic features. As a result, feature complementarity and redundancy are not explicitly considered, which can limit model efficiency and generalization. Neural Architecture Search (NAS) provides a promising approach to automatic architecture optimization, but its application to large heterogeneous search spaces remains computationally expensive and slow to converge. In this work, we propose a complementarity-guided monte carlo tree search with evolutionary fine-tuning framework (CG-MCTS-NAS). The framework defines a structured search space over five heterogeneous traffic features, including packet length, inter-arrival time, raw bytes, byte entropy, and flow-level statistics. A complementarity matrix derived from feature correlations is incorporated into the PUCT strategy of MCTS, biasing the search toward more complementary feature combinations and reducing redundant exploration. Based on this design, a two-stage optimization strategy is employed, combining global architecture search with evolutionary fine-tuning of hyperparameters. Experiments on multiple real-world datasets show that CG-MCTS-NAS consistently outperforms recent state-of-the-art methods in classification performance, while reducing the required MCTS search iterations by approximately 55.6\%. The complete code for the proposed method is available at https://github.com/yadonglu7-hue/CG-MCTS-NAS.

Automatic segmentation methods for lymphoma based on whole-body positron emission tomography (PET) and computed tomography (CT) still face significant challenges due to metabolic overlap between physiologically highly metabolizing and excreting organs (sFEPUs) and lesions. To address this, this paper proposes a whole-body lymphoma segmentation method that integrates anatomical and metabolic statistical priors. First, the CT-based whole-body organ segmentation method TotalSegmentator is employed to automatically segment 107 tissue/organ types across 300 PET/CT datasets. The metabolic intensity distribution of each tissue/organ is statistically analyzed to construct a population-level human metabolic statistical model. Building upon this foundation, an organ-specific metabolic suppression strategy is introduced for physiologically hypermetabolic regions, thereby reducing the complexity of the whole-body lymphoma segmentation task. Furthermore, a two-stage global-to-local segmentation framework is established: nnUNet performs coarse segmentation across the entire body, followed by local PET/CT-based fine segmentation using nnUNet to refine segmentation based on the probability map generated in the first stage, which locates potential lesion regions. Experimental results on the AutoPET dataset demonstrate that this method effectively reduces false positives caused by physiologically hypermetabolic organs and outperforms state-of-the-art medical image segmentation methods in overall segmentation performance.

The performance of medical image segmentation models faces the core challenge of data heterogeneity in multi-center collaborative training. Existing methods typically process mixed multi-center data directly while ignoring center-specific characteristics, resulting in suboptimal performance. To address this issue, this study proposes a Multi-Center Adaptive U-Net (MCA-UNet) for PET/CT images. The proposed method introduces a hierarchical center adaptation mechanism: center-specific adapters are employed in shallow layers to handle low-level variations, while multi-branch cross-center blocks in deeper layers balance general features and center-specific representations. In addition, MCA-UNet incorporates a low-rank dual-path fusion strategy to further enhance overall performance. Extensive experiments were conducted on the publicly available HECKTOR 2021 head and neck tumor segmentation dataset. Compared with other methods, MCA-UNet achieves superior performance, with a Dice score of 81.50% and an HD95 of 5.48 mm, with limited parameter growth (less than 5%). These results demonstrate that the proposed method can significantly improve overall segmentation performance on multi-center PET/CT images.

Software vulnerability detection is a critical task in cybersecurity. Existing hybrid methods combining Graph Neural Networks (GNNs) and Pre-trained Code Models (PCMs) face two fundamental limitations: structural noise in Code Property Graphs (CPGs) introduces spurious correlations, and the granularity mismatch between global semantics and local features results in insufficient fusion. To address these issues, we propose CausalVul, a robust framework based on dual-granularity semantic fusion and causally-motivated structural pruning. First, we design a curriculum learning-driven dynamic edge pruning mechanism that gradually shifts focus from global exploration to vulnerability-relevant causal subgraphs. Second, we construct a dual-granularity encoding architecture to extract function-level sequence semantics and fine-grained graph structural features in parallel, facilitating deep cross-modal fusion via gating mechanisms. Finally, we propose a multi-view optimization objective combining node-level Multiple Instance Learning, class center regularization, and supervised contrastive learning to enhance discriminative power on imbalanced data. Experiments on three widely used benchmarks (Devign, ReVeal, and Big-Vul) demonstrate that CausalVul achieves F1-scores of 65.05%, 51.01%, and 54.55%, respectively. Notably, on the highly imbalanced Big-Vul dataset, our method achieves competitive performance, demonstrating the effectiveness of the proposed approach on challenging real-world scenarios.

Evaluating cross-language code translation remains a fundamental challenge due to the inherent diversity of functionally equivalent implementations. Traditional reference-based metrics fail to account for valid solutions that diverge in syntax, while execution-based evaluation is often difficult to scale because it depends on runnable environments, test cases, and external dependencies. We propose CL-ICE (Cross-Language Internal Concept Evaluation), a white-box framework that probes the internal representations of Code Large Language Models (Code LLMs) to assess translation quality without execution or reference comparison at inference time. Our approach leverages two key insights from recent interpretability research: (1) intermediate layers of Code LLMs form language-agnostic concept layers—where abstract, universal representations of algorithmic logic are encoded regardless of surface syntax, and (2) LLMs encode an internal “correctness” signal in their activation space that can be extracted through Linear Artificial Tomography (LAT). CL-ICE includes two complementary components: concept-layer profiling for semantic invariance and language-specific LAT for latent correctness probing. Experiments demonstrate that CL-ICE achieves correlation with functional correctness (r = 0.582) that approaches that of GPT-4-based oracle evaluation (r = 0.615), while maintaining deterministic operation and low computational cost.

Large Language Models (LLMs) have demonstrated impressive capabilities in automated unit test generation. How- ever, they predominantly rely on textual co-occurrence patterns, often neglecting the explicit control flow and dependency struc- tures inherent in source code. This limitation frequently results in test cases that suffer from logical hallucinations and insufficient boundary coverage. While Graph Neural Networks (GNNs) excel at modeling non-Euclidean code structures, effectively aligning structural features with the textual semantics of LLMs remains a significant challenge. To address these issues, we propose SAD- Gen, a Structure-Aware Disentangled Generation framework. First, we present a dual-modal encoding architecture that lever- ages a textual encoder to capture sequential semantics alongside Graph Attention Networks (GAT) for extracting structural fea- tures from Abstract Syntax Trees (ASTs). These representations are dynamically fused via an adaptive gated mechanism. Second, to maximize generation diversity, we propose a novel Latent Space Disentanglement Strategy that decouples the program rep- resentation into immutable Logic Latents and perturbable Data Latents. By injecting Gaussian noise into the data latent space during generation, our model actively explores diverse boundary conditions while strictly preserving the underlying program logic. Extensive experiments on the Methods2Test and MBPP datasets demonstrate that SAD-Gen consistently outperforms state-of-the- art baselines in terms of syntax pass rate, branch coverage, and mutation score, validating the effectiveness and robustness of our proposed approach.

We present Postmortem Action Generation (PAG), a novel NLP framework for intelligently generating structured Corrective and Preventive Measures (CAPMs) from post-incident analyses in large-scale cloud systems. Unlike prior Root Cause Analysis (RCA) work that identifies why failures occur, PAG focuses on what to do next producing actionable, role-specific remediation steps. Our approach combines hybrid retrieval of similar incidents, persona-conditioned prompting, and LLM-based validation to ensure correctness, actionability, and role fit. Evaluated on over 8,000 real-world cloud incidents, PAG achieves strong performance with factual consistency (0.83), groundedness (85.3%), and expert-rated correctness (4.3/5), surpassing fine-tuned and rule-based baselines across automated and human evaluations. The architecture is lightweight, scalable, and production-ready frozen LLMs paired with retrieval and validation loops delivering reliable, large-scale post-incident decision support.

As information systems become increasingly complex, vulnerability detection has emerged as a critical means of identifying and mitigating security risks. Artificial intelligence-driven methods offer unique advantages in terms of automation and intelligence. However, the training process of current intelligent detection models tends to learn shallow patterns from the training data, overlooking the underlying semantic features of vulnerabilities. To address this challenge, we propose a novel vulnerability detection method named CoGRVD, which employs semantic-sensitive contrastive learning and token-level graph representation learning to capture critical semantic and structural features. Specifically, we design semantically equivalent yet vulnerability-aware code transformation methods that guide the model to focus on capturing meaningful vulnerability-related semantics. In addition, we transform the input into a token-level co-occurrence graph and employ GCNs with residual connection to effectively capture the long-term dependencies among tokens. We fuse the semantic and structural features and employ a classifier for prediction. To evaluate the effectiveness of our proposed method, we conduct extensive experiments on three public benchmark datasets. The experimental results demonstrate that CoGRVD outperforms all other baseline methods across all three datasets. Specifically, on the FFMPeg+Qemu, Reveal, and Big_vul datasets, its F1 scores reached 74.55%, 63.73%, and 90.23%, respectively, showing improvements of 7.61%, 13.16%, and 40.78% compared to the best baseline methods.

In software engineering, establishing accurate traceability links between requirements and source code is essential for many development and maintenance tasks, yet it remains challenging in large and structurally complex software systems. Recent retrieval-augmented generation approaches based on large language models have introduced new opportunities for automated traceability link recovery. However, their effectiveness is often constrained by retrieval quality and the vulnerability of single-pass decision mechanisms to reasoning errors and hallucinations. This paper proposes SARV, a framework that explicitly decouples candidate link generation from link validation. In the retrieval stage, a structure-aware hybrid retrieval strategy is employed to identify potential links. In the validation stage, SARV introduces a multi-agent collaborative mechanism guided by auxiliary information, which applies a reflect-and-correct strategy to transform traceability decisions from a one-shot classification process into a verifiable and error-correctable validation process. Experimental results on multiple public datasets show that SARV achieves competitive and stable performance across diverse project settings.

Current fault localization methods struggle to prioritize true faulty lines due to coverage ambiguities in dynamic analysis and missing semantic dependencies in static analysis. Existing feature-concatenation strategies fail to produce sufficiently discriminative suspiciousness scores for Top-N recall. To address this, we propose DSLA, which constructs a line-level property graph integrating dynamic coverage with static code semantics, enabling fine-grained code representation. DSLA performs deep feature fusion via a joint objective combining MSE, t-SNE constraints, and discriminative losses, explicitly enforcing geometric separability to prevent model collapse. On the Defects4J benchmark, DSLA improves Top-1 recall by 1.3× over state-of-the-art techniques, and on the C-language Codeflaws dataset, it achieves a 6.1× improvement, demonstrating its effectiveness in precise faulty line localization.

Large language models (LLMs) are increasingly effective at generating code, but they still struggle on specification-intensive tasks where reliability is critical. Such tasks typically come with long problem statements, layered constraints, and many corner cases, and they are often judged by hidden tests. To address this gap, we introduce DeepDebateCoder, a closed-loop framework designed to improve robustness and structural consistency in LLM-generated code. DeepDebateCoder first builds an explicit Policy Graph that captures module boundaries, branching logic, state transitions, and I/O constraints. This graph then serves as a structured guide for code synthesis and refinement. To go beyond example-level correctness, it generates policy-aligned tests that target boundary conditions and exception paths, and validates them before using them for refinement. Finally, DeepDebateCoder uses execution feedback to drive an asymmetric debate between a strategy expert and a code expert. Instead of repeatedly patching surface-level errors, the debate focuses on revising the underlying decision path. Across HumanEval and APPS, DeepDebateCoder consistently improves pass@1 and robustness over competitive baselines. The gains are especially pronounced on APPS, with the largest improvements on long-specification and high-difficulty tasks, underscoring the value of meta-policy modeling and few-shot-anchored asymmetric debate for achieving highly reliable code generation. More broadly, our framework points to a new paradigm for reliability-oriented program synthesis.

Smart contract vulnerability detection (SCVD) has attracted increasing attention in both academia and industry. Pre-trained language models (PLMs) are trained on large-scale data, have shown strong generalization ability. However, existing PLMs-based SCVD methods are typically limited to processing inputs of 512 or 1024 tokens, or lack structural information, which degrades vulnerability detection performance. An important question arises: How effective are existing PLMs when applied to SCVD? In this paper, we introduce a new multimodal feature-based vulnerability detection framework named MFVD. We utilize segmentation encoding strategy and graph presentation to tackle the limitations above. MFVD enhances UniXcoder to process multimodal data—the source code sequence and the data flow graph—and enables binary classification. Extensive experiments demonstrate that MFVD outperforms state-of-the-art methods in recall, precision, and F1 scores, as well as in the stability of these metrics, achieving 96.14% F1 for reentrancy vulnerabilities and 96.71% F1 for timestamp dependency vulnerabilities. We also evaluate 11 PLMs on source code, with results indicating that UniXcoder outperforms the others in most scenarios.

Machine learning is widely applied in software engineering tasks, particularly code vulnerability detection, owing to its ability to automatically identify patterns in data. However, existing machine-learning-based vulnerability detection methods exhibit several limitations. Reliance on a single feature type fails to capture comprehensive code semantics, leading to elevated false positive and false negative rates. Excessive dependence on fully automated feature learning overlooks potential inter-feature relationships, thereby impeding the model's capacity to recognize specific vulnerability patterns. Furthermore, the use of excessively high-dimensional features incurs substantial computational overhead during large-scale analysis, rendering real-time detection impractical. To address these shortcomings, we propose BAVul, a vulnerability detection system based on a bug attention mechanism. BAVul embeds source code using CodeBert language model to preserve vulnerability-relevant semantic information in the sequence. It then reconstructs the Control Flow Graph (CFG), Abstract Syntax Tree (AST), and Program Dependence Graph (PDG), and generates embeddings for these graphs via Graph Attention Network (GAT) to preserve structural information. During feature fusion, a bug attention mechanism is introduced to enable weighted integration of graph embeddings, thereby compensating for the limited modeling of vulnerability patterns in conventional graph-based approaches. In the detection phase, a refinement-based dimensionality reduction technique is combined with LightGBM to enhance both efficiency and accuracy. Compared with conventional machine-learning methods, BAVul achieves substantial improvements in F1-score and computational performance.

Early failure prediction from system logs is notoriously difficult because pre-failure signals are weak, heterogeneous, and easily overwhelmed by massive noisy logs. A key but under-exploited fact is that these early signals are structured across two levels: (i) rare semantic cues embedded in individual log messages, and (ii) temporal dynamics reflecting how the system state gradually drifts and occasionally transitions toward failure. Motivated by this observation, we propose a dual-level semantic and temporal feature augmentation approach for robust early failure prediction from noisy logs. At the semantic level, we introduce a random-walk-driven word reweighting mechanism that injects frequency-imbalance priors between normal and failure-related corpora, amplifying failure-indicative tokens without relying on log parsing or domain knowledge. At the temporal level, we design a complementary dynamics fusion predictor that couples an attention-based BiLSTM with multiple HMM components, where BiLSTM captures long-range non-linear evolution trends while HMMs model local Markovian state transitions associated with abnormal shocks. The two representations are then fused for accurate sequence classification. Extensive experiments on four public log datasets show that our method consistently outperforms strong failure prediction baselines and substantially surpasses representative log anomaly detectors, yielding higher F1 scores and improved recall under noisy conditions with practical inference overhead. These results demonstrate that jointly augmenting semantic saliency and temporal dynamics is crucial for reliable early failure prediction in real-world logging environments.

Code decompilation analysis is a fundamental yet challenging task in malware reverse engineering, particularly due to the pervasive use of sophisticated obfuscation techniques. Although recent large language models (LLMs) have shown promise in translating low-level representations into high-level source code, most existing approaches rely on generic code pretraining and lack adaptation to malicious software. We propose LLM4CodeRE, a domain-adaptive LLM framework for bidirectional code reverse engineering that supports both assembly-to-source decompilation and source-to-assembly translation within a unified model. To enable effective task adaptation, we introduce two complementary fine-tuning strategies: (i) a Multi-Adapter approach for task-specific syntactic and semantic alignment, and (ii) a Seq2Seq Unified approach using task-conditioned prefixes to enforce end-to-end generation constraints. Experimental results demonstrate that LLM4CodeRE outperforms existing decompilation tools and general-purpose code models, achieving robust bidirectional generalization.

Small Transformers trained from scratch on synthetic meta-learning corpora can learn to perform in-context learning (ICL), but doing so is notoriously slow and data-inefficient. We ask whether soft architectural priors-implemented as gated, learnable additions to attention logits can steer these tiny models toward ICL solutions faster while preserving flexibility. We augment attention heads with (i) a learnable relative-position prior and (ii) a content-matching prior, both controlled by learnable scalar gates regularised with weight decay so they can shrink toward zero once ICL circuitry forms. We evaluate two-layer Transformers on linear regression, associative recall, and algorithmic reasoning, comparing SOFT-BIAS against strong baselines including ALIBI and Learned Relative Positional Encodings (RPE). Our results show that SOFT-BIAS acts as an efficient catalyst: it accelerates convergence by approximately 2x compared to the VANILLA baseline. Crucially, in noisy regression settings where standard Learned RPE overfits, SOFT-BIAS remains robust due to its gating mechanism. Furthermore, on associative recall tasks, it achieves a superior trade-off between accuracy and training throughput compared to data-adaptive baselines. Analysis of learned gates confirms that the priors accelerate the emergence of ICL circuitry without constraining the model's final plasticity.

Photoplethysmography (PPG)-based blood pressure (BP) estimation is a complex biosignal processing task, particularly challenging for resource-constrained wearable devices. Deep learning methods have achieved state-of-the-art performance on large public datasets but still require hundreds of kilobytes of memory for deployment, even after strong optimization. In this work, we show the application of the spiking neural networks (SNNs). Thanks to the inherent suitability of SNNs for temporal signals, our best quantized network achieves comparable or better performance on four benchmark datasets compared to state-of-the-art (SOTA) deep neural networks (DNNs), while strongly reducing the number of parameters by up to 13.8$\times$. We also demonstrate that our optimized SNN can be executed on wearable devices, deploying it on a low-power RISC-V-based microcontroller unit (MCU) and obtaining a memory footprint of 4.63 kB, a latency of 1.34 ms, and an energy consumption of 0.068 mJ.

Convolution is a principal computational bottleneck in deep neural networks, and its efficiency depends on tight integration between algorithms and GPU hardware. Existing GPU convolution methods suffer from large memory overhead, poor cache utilization, limited effectiveness across kernel sizes, or numerical instability. This work extends the im2win paradigm—a universal, memory-efficient convolution method with contiguous memory access for all kernel sizes—to run efficiently in full precision on CUDA cores and half precision on tensor cores. By introducing new kernel designs and optimizations such as zig-zag memory access and asynchronous data movement, im2win efficiently exploits hardware-accelerated half-precision matrix multiply-accumulate operations. Across twelve CNN benchmarks, im2win achieves up to 2.8$\times$ higher TFLOPS than its CUDA core implementation, 1.4$\times$ higher than cuDNN, and 6.4$\times$ higher than GEMM-based convolution with cuBLAS, while using as little as 53\% and 35\% of their memory, respectively. These results establish im2win as a unified, high-performance convolution framework for modern GPU architectures.

Wi-Fi sensing on resource-constrained platforms calls for efficient and tiny neural networks, yet CSI pipelines face a recurring trade-off between preserving the full multi-antenna CSI tensor and compressing the input by simple heuristics. We propose ST-CA2 (SpatioTemporal Channel–Antenna Attention), a lightweight input refinement module that performs subcarrier wise channel attention and antenna-pair-wise attention to output a compact C-by-T representation. Under a fixed downstream recognizer and in a door-opening person-identification setting, ST-CA2 achieves 99.20% OA with 6.208 GFLOPs and 2.006 GB peak GPU memory, compared with 97.20% OA with 53.232 GFLOPs and 8.601 GB for full-input. These results indicate a favorable accuracy–efficiency trade-off for compact CSI input refinement in the studied setting. We also provide analyses under signal-plus-noise assumptions and evaluate temporal statistical priors (GTA, variance, and RMS) for weight generation.

Early diagnosis of Autism Spectrum Disorder (ASD) plays a crucial role in improving patients’ quality of life. In recent years, face image–based ASD detection has attracted increasing attention as an auxiliary diagnostic approach. However, existing lightweight models still show limitations in capturing fine-grained facial features. To address this problem, this paper proposes a lightweight face image–based auxiliary detection model for ASD, termed MN-ASD. First, MobileNetV4-S is selected as the baseline framework. To enhance the model’s ability to capture subtle facial details, the FReLU activation function is introduced, which strengthens spatial feature modeling. Second, to further improve the performance and accuracy of the model, the Coordinate Attention (CA) module is incorporated at the final stage. By jointly modeling positional and channel dependencies, the CA mechanism enables the model to better locate key facial regions and enhance global feature extraction. Finally, experimental results demonstrate that the proposed method achieves an accuracy of 93.33% on the Kaggle dataset, with precision, recall, and F1-score all outperforming other mainstream lightweight networks. These results demonstrate the effectiveness of the proposed MN-ASD model in capturing discriminative facial features, while highlighting its strong potential for deployment in resource constrained and TinyML-oriented environments.

The Lottery Ticket Hypothesis for deep neural networks emphasizes the importance of initialization used to retrain the sparser networks obtained using the iterative magnitude pruning process. An explanation for why the specific initialization proposed by the lottery ticket hypothesis tends to work better in terms of generalization (and training) performance has been lacking. Moreover, the underlying principles in iterative magnitude pruning, like the pruning of smaller magnitude weights and the role of the iterative process, lack full understanding and explanation. In this work, we attempt to provide insights into these phenomena by empirically studying the volume/geometry and loss landscape characteristics of the solutions obtained at various stages of the iterative magnitude pruning process.

Quantization-aware training (QAT) and Knowledge Distillation (KD) are combined to achieve competitive performance in creating low-bit deep learning models. Existing KD and QAT works focus on improving the accuracy of quantized models from the network output perspective by designing better KD loss functions or optimizing QAT’s forward and backward propagation. However, limited attention has been given to understanding the impact of input transformations, such as data augmentation (DA). The relationship between quantization-aware KD and DA remains unexplored. In this paper, we address the question: how to select a good DA in quantization-aware KD, especially for models with low precision? We propose a novel metric which evaluates DAs according to their capacity to maximize the Generalized Contextual Mutual Information--the information not directly related to an image's labels--while also ensuring the predictions for each class are close to the ground truth labels on average. The proposed method automatically ranks and selects DAs, requiring minimal training overhead, and it is compatible with any KD or QAT algorithm. Extensive evaluations demonstrate that selecting DA strategies using our metric significantly improves state-of-the-art QAT and KD works across various model architectures and datasets.

Speculative decoding (SD) accelerates large language model (LLM) inference through introducing a fast draft model to propose candidate token sequences. However, conventional draft models employ softmax-based attention for capturing original transformers, incurring quadratic time and memory complexity during prefilling. Given the limited resources assigned to drafters, this design restricts their scalability in long-context scenarios. To this end, we propose a new drafter architecture built on linear attention, which achieves linear computational complexity. Specifically, our method maintains a fixed-size key-value cache during autoregressive generation, enabling efficient, scalable inference with minimal overhead. We evaluated the proposed approach on multiple model scales and benchmarks, demonstrating comparable token acceptance rates and downstream performance to standard attention mechanisms, while achieving an average 2.62 times speed-up.

We introduce FedTinyProp, a Federated Learning (FL) framework that adapts sparse backpropagation for on-device training. In our approach, clients (edge devices) dynamically sparsify gradients based on per-batch informativeness and skip updates when computation is unnecessary. Clients transmit sparsified weight deltas that are aggregated via a communication-aware Federated Averaging variant.

Experiments across FashionMNIST, CIFAR-10/100, UCI HAR, and Google Speech Commands show that FedTinyProp maintains accuracy within 1 to 6 points of dense baselines while reducing training compute by about 2--5x via adaptive sparsity and skipping. In addition, it reduces client upload volume by up to about 65% and lowers peak activation/gradient memory by up to about 45% through sparse backpropagation and sparse delta exchange. Despite the added work of sparse selection, controller overhead is negligible and the savings-to-overhead trade-off remains favorable, with large FLOP savings outweighing non-backpropagation costs.

Human activity recognition (HAR) using wearable sensors has gained significant attention for healthcare monitoring and smart device applications. However, deploying deep learning models on resource-constrained edge devices remains challenging due to high computational costs and memory requirements. In this paper, we propose LiteTSC, an ultra-lightweight neural network for time series classification that achieves competitive accuracy with significantly reduced model complexity. Our approach integrates three key components: (1) multi-scale depthwise separable convolutions that capture temporal patterns at different granularities while reducing parameters, (2) a lightweight dual-pooling channel attention mechanism that enhances feature discrimination with minimal overhead, and (3) residual connections that compensate for the reduced representational capacity of lightweight operations. Extensive experiments on three benchmark datasets (UCI HAR, WISDM, and PAMAP2) demonstrate that LiteTSC achieves state-of-the-art accuracy (96.74% on UCI HAR) with only 20.8K parameters. Compared to TinyHAR, the ISWC 2022 Best Paper, LiteTSC achieves higher accuracy with 14.3× fewer parameters and 3.3× faster inference. The lightweight design makes LiteTSC suitable for deployment on resource-constrained devices.

Atmospheric vortices on Mars, such as dust devils, play a critical role in the planet’s climate, dust transport, and surface--atmosphere interactions. Capturing high-rate wind, pressure, and dust sensor data during vortex encounters provides valuable scientific insight; however, such events are rare and unpredictable. Continuous high-rate data collection is therefore impractical due to strict power and onboard data storage constraints, particularly for low-cost landers. Prior work has proposed using machine learning algorithms to detect approaching vortices from low-rate pressure measurements and autonomously trigger high-rate data acquisition during the most scientifically relevant portion of the event. In this work, we extend prior vortex detection studies by evaluating both classical machine learning models and a two-stage deep learning pipeline for early-warning detection under precision-critical deployment constraints. Specifically, we assess Random Forest and Extreme Gradient Boosting (XGBoost) classifiers trained on statistically engineered features, alongside a two-stage autoencoder-gated Temporal Convolutional Network (AE$\rightarrow$TCN) designed to filter nominal atmospheric behavior prior to classification. Model performance is evaluated using metrics based on precision, recall, and F1-score appropriate for rare-event detection. In addition, we examine the feasibility of onboard deployment by benchmarking inference latency and power consumption on a Qualcomm Snapdragon–class processor representative of spaceflight and edge-computing hardware. By jointly analyzing detection performance and computational efficiency, this study provides a comparative assessment of accuracy–resource trade-offs relevant to autonomous onboard triggering and demonstrates the potential of learning-enabled intelligent sensing to enhance Martian atmospheric science while respecting mission resource constraints.

Standard video action recognition models often process typically resized full frames, suffering from spatial redundancy and high computational costs. To address this, we introduce MoCrop, a motion-aware adaptive cropping module designed for efficient video action recognition in the compressed domain. Leveraging Motion Vectors (MVs) naturally available in H.264 video, MoCrop localizes motion-dense regions to produce adaptive crops at inference without requiring any training or parameter updates. Our lightweight pipeline synergizes three key components: Merge & Denoise (MD) for outlier filtering, Monte Carlo Sampling (MCS) for efficient importance sampling, and Motion Grid Search (MGS) for optimal region localization. This design allows MoCrop to serve as a versatile "plug-and-play" module for diverse backbones. Extensive experiments on UCF101 demonstrate that MoCrop serves as both an accelerator and an enhancer. With ResNet-50, it achieves a +3.5% boost in Top-1 accuracy at equivalent FLOPs (Attention Setting), or a +2.4% accuracy gain with 26.5% fewer FLOPs (Efficiency Setting). When applied to CoViAR, it improves accuracy to 89.2% or reduces computation by roughly 27% (from 11.6 to 8.5 GFLOPs). Consistent gains across MobileNet-V3, EfficientNet-B1, and Swin-B confirm its strong generality and suitability for real-time deployment. Our code and models are available at https://github.com/microa/MoCrop.

Standard evolutionary search algorithms, including differential evolution (DE), use a relatively small, limited-size population that remains the same size or gradually shrinks during the search. It has long been believed that replacement (discarding some members of the population and inserting new candidate individuals in their place) is an important feature of evolutionary search. Recent work on Unbounded DE (UDE) has shown that non-increasing populations are not necessary, and that with an appropriate selection operator, a DE with a monotonically growing population can perform comparably to state-of-the-art adaptive DE. However, UDE has scalability issues, as the runtime complexity of each iteration increases with population size, and limited RAM capacity poses a limitation on the growth of the population. We propose extensions to UDE that overcome these limitations by (1) efficiently storing the population on external storage (SSD) and (2) reducing the computational complexity of tournament selection. We show experimentally that this allows UDE to scale to very large populations with 2e8 individuals.

Test case prioritization (TCP) addresses the challenge of efficiently ordering test cases to improve fault detection rates while managing execution time constraints in regression testing. This study presents HADI framework for multi-objective test case prioritization using Non-dominated Sorting Genetic Algorithm II (NSGA-II). The study evaluates the framework using the Defects4J benchmark. Our framework optimizes two competing objectives: (i) minimizing the execution time to fault detection (ETAF), and (ii) maximizing the Average Percentage of Faults Detected (APFD). The findings show that HADI achieves a mean APFD of 0.98 compared to 0.51 for original orderings, representing a 92% improvement. Statistical validation using Wilcoxon signed-rank tests confirms significance (p < 0.001) with large effect sizes (Vargha-Delaney A12 > 0.89). Convergence analysis shows that HADI typically reaches optimal solutions within 10-20 generations, with consistent performance across 30 independent runs per dataset. Index Terms—Test case prioritization, NSGA-II, multiobjective optimization, Defects4J, Pareto optimization.

The ϵ-constraint handling mechanism is a widely used method for dealing with constraints. However, when the value of ϵ is reduced to a very small magnitude, it becomes equivalent to the constrained dominance principle. Under these circumstances, maintaining the diversity of the algorithm becomes challenging. To overcome this challenge, we introduce AEP-PPS, an adaptive epsilon-penalty switching for push-pull search algorithm with the adaptive stage transition mechanism. AEP-PPS employs a two-stage search strategy. In the push stage, constraints are temporarily ignored to drive the population toward the UPF. Based on the symmetric cross-generational distance indicator, the algorithm then adaptively switches to the pull stage, in which an ϵ-constraint handling mechanism combined with a penalty function is adopted to progressively enhance feasibility pressure and guide the population toward the CPF. In addition, an archive updating strategy based on θ-domination is introduced to preserve population diversity. Experimental results on the LIR-CMOP and CF test suites demonstrate that AEP-PPS outperforms nine state-of-the-art algorithms in terms of IGD and HV metrics. These results confirm its effectiveness in handling complex infeasible regions while achieving a good balance between convergence and diversity.

This paper presents a novel memetic hyperspectral unmixing approach based on the MOEA/D framework, focusing on endmember identification using a spectral library. The population is optimized by alternately applying genetic and local search operators. The local search operator is designed to extract and utilize knowledge accumulated across iterations, particularly the endmember abundances of each individual. This enables it to guide individual optimization toward the direction dictated by cultural genes. Additionally, a voting mechanism integrates multiple Pareto-optimal solutions by selecting the most frequently occurring endmembers as the final set, thereby mitigating the instability inherent in relying on a single solution. Extensive experiments on simulated and real datasets demonstrate that incorporating cultural genes significantly enhances the evolutionary algorithm’s performance, yielding superior unmixing results.

To address the limited adaptability of existing psychological dialogue data in the field of adolescent mental health support, and the insufficient guidance ability of current models, the Adolescent Psychological Support Dialogue Generation Model of Domain Data-Driven (APS-DM) is proposed. By extracting multi-level adolescent psychological topics, using topic-related keywords to filter and integrate data corpus, a context-based role-playing prompt is designed according to the characteristics of adolescent psychological development stages. It is intended to guide GPT-4 in performing multi-turn semantic reconstruction, thereby constructing an adolescent psychological support dialogue dataset (AdoPsyDialogue). Besides, a structured thinking-guided prompt is designed to constrain the response generation scenarios. A parameter-efficient fine-tuning method is used to optimize the base model, enhancing its ability to provide adolescent psychological support responses and guidance. Experimental results on the QsnTest dataset demonstrate that the APS-DM model significantly outperforms models such as ChatGLM2-6B and ChatGPT-3.5 in both automatic and manual metrics.

Patients with chronic kidney disease (CKD) undergoing peritoneal dialysis (PD) commonly experience multimorbidity and polypharmacy, making adverse drug reaction (ADR) risk highly patient-specific. Conventional trials and drug-centric computational models often lack sufficient patient context and longitudinal validity to capture context-specific safety signals in real-world PD follow-ups. We propose PS-PDNet, a patient-state–conditioned deep learning framework that integrates clinical indicators and longitudinal prescriptions to predict ADR risk in PD. By tailoring pharmacodynamics modeling to each patient’s clinical context and temporally valid exposure history, PS-PDNet supports individualized safety assessment. Coupled with counterfactual inference, it highlights high-risk drugs and combinations and summarizes them into clinically interpretable category-level exposure patterns, providing actionable cues for ADR surveillance and medication optimization in real-world PD care.

Deep learning models for medical image classification often underperform when class imbalance is extreme, as supervised methods overfit the majority class, leading to biased feature representations. To address this, we propose Self-Calibrated Feature Denoising (SCFD), a framework that enhances the performance of self-supervised models in imbalanced settings. SCFD integrates a pretrained DINOv2 encoder for class-specific feature extraction, a lightweight autoencoder to denoise minority-class embeddings, and a latent-space regularization mechanism that leverages majority-class information for calibration. By combining these components, SCFD generates balanced, discriminative embeddings without modifying the backbone, enabling seamless integration into existing pipelines. Experiments on three disease diagnosis datasets, Lab-shanghai, Brain Tumor 4C, and Chest CT, demonstrate that SCFD consistently improves model robustness and generalization under extreme class imbalance, providing a reliable approach for effective medical image classification.

Breast cancer histopathological image analysis is crucial for diagnosis and treatment planning. Although multiple instance learning (MIL) methods based on whole slide images (WSIs) have shown great promise, existing approaches often suffer from two limitations: overfitting caused by insufficient training samples and a training bias dominated by easily classified instances, which leads to the neglect of hard examples. To address these issues, we propose a multiple instance learning framework for breast cancer based on pseudo-bag augmentation and double-layer masking (PADM-MIL). Specifically, PADM-MIL first expands the training set via pseudo-bag partitioning and mixing. It then integrates a hierarchical Transformer and introduces a teacher network to construct slide-level prototypes. Interpretable instance importance (importance score) is computed using the cosine similarity between instance representations and prototypes. These importance scores further guide instance-level masking to mine hard examples. Subsequently, the student network performs additional group-level masking at the sub-bag level on hard-example bags, enabling the model to stably focus on key tissue regions. Experiments on two public breast cancer datasets demonstrate that PADM-MIL consistently outperforms existing methods in terms of AUC, ACC, and F1-score. Results on the TCGA-LUNG lung cancer dataset further indicate that the proposed framework exhibits promising generalization.

Medical image segmentation plays a crucial role in disease detection, diagnosis and treatment. However, with the rapid increase in three-dimensional (3D) image data size, existing medical segmentation methods struggle in balancing computational cost and model performance. This paper proposes a slice-based segmentation method on 3D images and several plug-and-play Slice Context Modules (SCMs) to enable 2D networks to perform 3D segmentation. With an autoregressive method to extract inter-slice context information, SCMs use image morphological operations to expand this information and predict the inter-slice morphological changes with previous inter-slice context information, and use a channel-wise fusion module to guide the segmentation. SCMs improve the model's ability to sense and predict target morphological changes, while balancing the accuracy and efficiency in model training process. This paper also designed a specific training method called random partitioning to further accelerate the training process. Experiments conducted on open datasets have shown that the proposed methods increased the performance of baseline networks by over 4.6% in Dice score, surpassing multiple state-of-the-art 3D segmentation networks. Furthermore, SCMs achieve 30% faster training speed and 40% less GPU memory usage during training, demonstrating their effectiveness in 3D medical segmentation tasks.

Medical image segmentation is a challenging task, where U-shaped networks are widely used. In order to better model long-range dependencies, many recent works integrate Transformer or state-space model (SSM) into U-Net and claim superior performance. However, such methods either suffer from semantic redundancy and heavy computation in self-attention, or require unnatural scanning routes in SSM. To validate their limitations, we first conduct an empirical study to show that existing Transformer- or SSM-based methods may actually fail to outperform vanilla U-Net under fair comparison settings, i.e., with the same pre -trained encoder and proper ablations. Then, to mitigate the limitations of self-attention, we propose BIdirectional Dynamic TransFormer (BIDFormer) to model spatial relationships and semantic correlations separately. (i) spatial relationships: to efficiently model global context, we replace self-attention with Bidirectional Linear Attention (BLA) between image patches and learnable class embeddings; (ii) semantic correlations: to explicitly model the correlation among target classes, we utilize class embeddings as Dynamic Low-rank Classifiers (DLC). Extensive experiments on public datasets demonstrate that BIDFormer can be easily integrated into existing models, leading to remarkable performance improvements. Code is available at https://github.com/JingTongsh/UN-Segmentation.

Current oral imaging modalities, such as panoramic radiographs and intraoral photographs, suffer from challenges including noise, blurring, occlusions, and significant variations in pose and scale. Intraoral photographs further exhibit specular reflections and complex illumination. Moreover, tooth arrangement is governed by strict morphological and anatomical constraints. These factors collectively pose significant challenges to accurate tooth position detection. To address these challenges, we propose a novel network architecture. Specifically, we introduce a Hybrid Frequency--Spatial Attention (HFSA) capable of simultaneously capturing both frequency and spatial features from images, thereby enhancing the network's robustness against various noise conditions. Furthermore, we design a Differentiable Class-Uniqueness Gating (UniGate) Mechanism to impose unique indexing constraints on each tooth. Additionally, we develop an Arch-aware Query Graph Prior (QGP), which enforces morphological and anatomical constraints on the detection process. We validate the performance of our approach through experiments on a public DSLR intraoral dataset, a private smartphone in-the-wild intraoral photograph dataset, and the public DENTEX2023 benchmark. Our method achieves state-of-the-art performance across all three datasets with AP scores of 78.3%, 54.4%, and 54.1%, respectively.

Tooth position detection underpins automated analysis of oral images, yet existing methods often fail on low-quality radiographs with blur, overlapping teeth, and low contrast. We propose GQD-DETR, an encoder–decoder framework tailored for robust tooth localization under real-world imaging conditions. GQD-DETR comprises three key components: (1) a Geometry-Aware Attentive High-Resolution module that combines polar coordinate priors with a high-resolution attention pathway to capture dental-arch curvature and fine details; (2) a Quality-Adaptive Cross-Scale Interaction module that performs quality-aware feature filtering and cross-scale context aggregation to handle noise and illumination inhomogeneity; and (3) a Discriminative Anisotropic Decoding module with learnable anisotropic positional encoding and query channel attention, which stabilizes query refinement in crowded regions and improves discrimination between spatially adjacent, visually similar teeth. Experiments on our private dataset and the public DENTEX2023 benchmark show that GQD-DETR achieves 58.98 mAP and 54.67 mAP, outperforming strong baselines by 1.62 and 1.43 mAP, respectively, and effectively handling multi-source images including panoramic radiographs, smartphone photos, and DSLR images.

Fine-tuningvision–languagemodels(VLMs)like CLIPformedicalscreeningfacestwocriticalchallenges:noisy supervisionfromweakly-alignedimage–textpairsandperfor-mancedisparitiesacrossdemographicsubgroups.Inglaucoma screening,wherefundusimagesarepairedwithcondensedclini-calnotes,theseissuesmanifestasheavy-tailedcontrastivelossdis-tributionsandunequalaccuracyacrossprotectedattributes.We proposeaunifedframeworkaddressingbothchallengesthrough threesynergisticcomponents:(1)TokenResidualFusion(TRF)strengthensvisualrepresentationsbyblendingintermediateand fnalViTtokenfeatures,(2)Student-tInfuenceReweighting(T-IRLS)stabilizesoptimizationbysuppressingextremeresiduals viarobuststatisticalweighting,and(3)Class-ConditionalCenter Alignment(CC-CA)reducessubgroupdisparitiesthroughen-tropicoptimaltransporttodynamicallyupdatedclasscenters.Ourmethodrequiresnoarchitecturalchangesatinference,addsminimaltrainingoverhead,anddemonstratesconsistent improvementsinbothoverallAUC(upto5.8%gain)andsub-groupfairnessacrossmultipleprotectedattributesinglaucoma screening.

Transformer-based models have achieved strong performance in single-cell transcriptomics, yet scaling them to genome-wide gene sequences remains challenging due to the quadratic cost of self-attention. Recent state space models (SSMs) offer linear-time sequence modeling, but unidirectional designs may introduce ordering biases and overlook biological structure, potentially limiting robustness across datasets. To address these challenges, we propose GeneMamba, a bidirectional state space model tailored for single-cell RNA sequencing. GeneMamba integrates three key components: (1) Efficiency: BiMamba blocks enable bidirectional processing with O(L) complexity, providing substantial speedup over transformer baselines while maintaining linear scaling; (2) Biological integration: a pathway-aware contrastive objective incorporates curated pathway information to guide representation learning; (3) Scalability: large-scale pretraining on approximately 30 million curated cells (from about 50 million raw CELLXGENE profiles) supports transfer across multiple downstream tasks without task-specific architectures. Across benchmarks including multi-batch integration, cell type annotation, and gene rank reconstruction, GeneMamba demonstrates competitive performance relative to strong baselines such as scGPT and Geneformer, with improvements observed in several settings. Ablation studies further suggest that bidirectional modeling and pathway-aware pretraining contribute to these gains. Overall, GeneMamba provides an efficient and biologically informed framework for foundation modeling in single-cell transcriptomics.

Diabetic Retinopathy (DR) is a leading cause of preventable blindness globally, yet early diagnosis faces signif- icant challenges, including a lack of annotated data and limited access to ophthalmic expertise. While machine learning based screening systems (e.g., Convolutional Neural Networks) show promise for DR prediction using retinal image analysis, they typically require a large number of labeled images, an expensive and time-intensive demand that is unsuitable for resource- constrained settings. Semi-supervised learning (SSL) offers an alternative by harnessing abundant unlabeled images alongside limited labeled data. However, existing SSL methods for DR detection struggle with unreliable pseudo-labels. We propose a novel SSL framework that adapts area under margin self-training (AUM-ST), originally developed for natural language processing, to medical imaging for DR detection. Our approach addresses unreliable pseudo-labelling by leveraging training dynamics to assess label quality. We replace the text-based BERT encoder with BEiT (Bidirectional Encoder representation from Image Transformers) to process retinal fundus images while preserving AUM-ST’s core reliability assessment mechanism. Additionally, we implement saliency-preserving augmentation strategies that avoid aggressive transformations such as random rotations and crops, thereby retaining critical lesion features essential for accu- rate diagnosis. Experimental validation on a real-world dataset of 757 retinal fundus images demonstrates that the proposed method outperforms established baselines, including ResNet- 18, ResNet-50, DenseNet-121, MobileNet-v2, and EfficientNet-B0. Our approach achieves substantial improvements in accuracy and F1 scores when training with fewer than 50 labeled samples per class, an advancement over traditional semi-supervised methods. The results demonstrate the practical viability of the proposed method for medical imaging applications and its potential for DR screening in resource-limited environments, including rural healthcare facilities and developing regions.

Early intervention is critical for ultrasound screening of developmental dysplasia of the hip (DDH), yet speckle noise, weak textures, and blurred boundaries often hinder reliable detection and accurate localization of anatomical landmarks. Although structures such as the cartilage–bone interface and the femoral head contour provide strong edge cues, DETR-style detectors typically emphasize region-level semantics and make limited use of explicit boundary and structural priors. Moreover, Transformer decoders usually rely on iterative refinement, which slows convergence and can degrade performance on small or crowded targets.We propose a structure-aware DETR framework tailored for DDH ultrasound that jointly performs object detection and six keypoint localization. An Edge-aware Feature Propagation (EFP) module extracts multi-scale edge representations from shallow features and strengthens boundary sensitivity via convolutional edge fusion. In addition, we introduce a Heatmap-guided Keypoint Prior (HKP) branch and a Heatmap-guided Query Patch Embedding (HQPE). A coarse keypoint heatmap head on P3 converts sparse keypoint supervision into six-channel Gaussian heatmap regression, providing dense spatial constraints and improved robustness; the features are further enhanced through heatmap-aligned referencing. Experiments on our in-house DDH ultrasound dataset (1,255 training images and 300 test images) demonstrate consistent gains over DETR, DeiM, and strong YOLO baselines in terms of mAP@0.75 and PCK, with notably more stable localization under noise and boundary ambiguity.

While question answering (QA) technologies have progressed significantly, their deployment for primary-school children remains understudied. Children’s queries often exhibit semantic ambiguity, incomplete expressions, and unclear articulation, leading to unanswered questions or irrelevant responses. Leveraging external knowledge is key to addressing this common challenge of ambiguity. Consequently, effectively reformulating questions and refining prompts into precise SPARQL queries are crucial tasks. This paper presents a multi-task learning approach utilizing a text generation model for both question reformulation and prompt refinement. Additionally, to bridge the preference gap between the question reformulation module and the QA model, we introduce a training strategy that utilizes QA model feedback to further optimize the question reformulation processes. Evaluations across two datasets show our method outperforms baselines, achieving statistically significant gains of +5.96% to +12.70% in F1 and Acc scores. These consistent improvements confirm that unified question reformulation and prompt refinement, reinforced by downstream QA feedback, substantially advance child-oriented conversational QA.

Knowledge base question answering (KBQA) aims to retrieve precise answers from knowledge graphs using natural language queries. While large language models (LLMs) enhance KBQA with strong reasoning capacities, they still face challenges such as noisy retrieval paths and inherent difficulties in reasoning over graph structures. To address these issues, we propose a novel framework, Path Retrieval and Deductive Reasoning (PRDR), which reformulates path retrieval as a binary classification task to suppress irrelevant paths and facilitate accurate reasoning. Retrieved paths are converted into propositional forms, enabling LLMs to perform structured deductive inference. Our method achieves state-of-the-art results on the WebQuestionsSP and ComplexWebQuestions benchmarks, demonstrating robust performance in complex multi-hop question answering.

Context Query Rewriting (CQR) is pivotal in conversational search for resolving context dependencies. However,existing methods typically employ a brute-force strategy that feeds the entire dialogue history into Large Language Models (LLMs). This indiscriminate approach introduces severe noise—particularly in multi-turn conversations with frequent topic shifts,which hinders both model robustness and training efficiency. To overcome these limitations, we propose CTS-CRL (Context Turn Selection for Conversational Rewriting with Joint Learning). Unlike previous methods, CTS-CRL incorporates a novel semantic-temporal context selection mechanism that dynamically filters irrelevant turns while preserving essential dependencies. Furthermore, we integrate this mechanism into a unified training framework combining Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). Experimental results on two benchmarks demonstrate the superiority of our approach. Notably, CTS-CRL establishes a new state-of-the-art on the TopiOCQA dataset, significantly outperforming baselines in handling complex topic transitions, while maintaining high efficiency with fewer parameters.

Inductive reasoning on Temporal Knowledge Graphs (TKGs) is pivotal for generalizing to entities unseen during training. However, existing symbolic methods heavily rely on historical interactions, rendering them ineffective under the extreme data sparsity of new entities (the "cold-start" problem). To address this, we propose Temporal Hierarchical Adaptive Inductive Logic (T-HAIL). First, we introduce an Open-to-Closed Knowledge Injection (OC-KI) mechanism, which leverages the parametric causal knowledge of Large Language Models (LLMs) to generate candidate rules, compensating for the lack of statistical signals in sparse graphs. To bridge the vocabulary gap, we employ a Semantic Vector Alignment module via Sentence-BERT to map candidates to the KG schema. These rules are refined through a rigorous data-grounding mechanism—validating candidates against observed facts—to eliminate hallucinations. To fully exploit these grounded rules under sparsity, we introduce a Hierarchical Adaptive Inference (HAI) mechanism that dynamically transitions between Exact Rule Matching, Semantic Similarity Borrowing, and Topology-based Fallback. Extensive experiments on ICEWS and GDELT benchmarks demonstrate that T-HAIL significantly outperforms state-of-the-art baselines in mitigating sparsity issues. Our code and datasets are available at: https://anonymous.4open.science/r/T-HAIL-B866.

Knowledge graph completion (KGC) aims to infer missing links by reasoning over triples. Recent work leverages Large Language Models (LLMs) to enhance graph understanding, yet three key challenges remain. First, LLMs often produce ambiguous entity descriptions lacking relational grounding. Second, relation semantics are context-dependent on the conditional entity, a factor frequently overlooked. Third, semantic embeddings from LLMs are misaligned with structural representations, limiting their effectiveness. In this paper, we propose ISERA-KGC (Entity–Relation Interactive Semantic Enhancement and Representation Alignment for Knowledge Graph Completion), a Language Model as a Service framework that leverages frozen LLMs to provide role-sensitive, context-aware semantic descriptions for KGC. Our method introduces two prompting strategies: Role-Aware Relation Prompting (RARP) for generating direction-specific relation semantics, and Relation-Anchored Entity Reasoning (RAER) for grounding ambiguous entities via relational context. Then the Contrastive Fusion and Alignment (CFA) module leverages contrastive learning to align and fuse these semantic and structural embeddings. Experiments on benchmark datasets FB15k-237 and WN18RR show that ISERA-KGC consistently improves mean reciprocal rank (MRR) by up to 2.4 points across multiple KGE models. These results demonstrate the effectiveness of our method in enhancing long-tail reasoning and bridging semantic-structural gaps in knowledge graph completion.

Large Language Models (LLMs) have demonstrated powerful performance in knowledge-intensive reasoning tasks. However, when dealing with large-scale Knowledge Graphs (KGs), they still struggle with hallucinations, inefficient multi-hop reasoning, and sensitivity to noisy retrieval results. Existing graph-based reasoning frameworks typically rely on fixed-hop search strategies, discrete decision boundaries, and the blind linearization of subgraph inputs, which often lead to excessive contextual noise and unnecessary computational overhead.

To address these issues, this paper proposes DRAoG, an adaptive dynamic graph reasoning framework capable of dynamically controlling reasoning depth and enforcing answer grounding during the generation process. The framework consists of three core components: relation-level subgraph filtering, reasoning state-aware decision model, and semantic matching-based grounding verification. Unlike training-based methods, DRAoG operates in a plug-and-play manner without requiring additional model fine-tuning.

Experiments on four public Knowledge Base Question Answering benchmarks demonstrate that the DRAoG framework achieves an accuracy improvement of 3.0% to 8.4% on multi-hop datasets while reducing LLM API calls by 15%. These results indicate that adaptive reasoning control and lightweight grounding verification provide an effective and practical solution for scalable graph-based LLM reasoning.

Event Argument Extraction (EAE) is essential for identifying event participants and their roles. However, in few-shot settings, existing generative approaches often depend on manually crafted, ontology-specific prompts and formats—an expensive and error-prone process that transfers poorly across domains and schemas. To mitigate these limitations, we propose FGEE, a universal JSON-based generation framework. FGEE tackles the data-scarcity bottleneck through three key innovations: (1) Unified I/O formats, which represent complex event structures with a standardized JSON schema and eliminate per-ontology prompt engineering; (2) Pseudo-output retrieval, which retrieves structurally similar pseudo-outputs to improve task understanding and mine salient event clues; and (3) Key sentence-based reasoning, which performs trigger-guided analysis to focus on informative evidence and deduce arguments from redundant contexts. We further introduce an N-Event-K-Shot evaluation framework to assess EAE under strict few-shot conditions. Experiments on ACE05, RAMS and WikiEvents datasets show that FGEE consistently outperforms the latest few-shot supervised models, yielding about a 9.6\% improvement in F1 across all benchmarks.These results show that FGEE substantially improves the robustness and generalization of generative EAE in low-resource environments.

Relation Extraction (RE) is a core task in information extraction. Although existing approaches have achieved substantial progress in modeling explicit relations, their performance remains limited when confronted with implicit relations that require contextual reasoning or commonsense knowledge, particularly in complex scenarios involving multiple entities and multiple relations. To address this challenge, this paper proposes Implicit Relation Extraction with Large Language Models (IMP-RELLM). The proposed approach explicitly introduces implicit information generated by Large Language Models (LLMs) to enhance the model’s ability to capture deep semantic representations and latent relational patterns in text. Specifically, IMP-RELLM adopts an instruction fine-tuning paradigm that organizes task descriptions, predefined relation sets, implicit information, and the original context into a unified structured prompt, guiding the model to attend to semantically valid entity relations that are not explicitly expressed in the text during generation. To reduce training costs and improve scalability, parameter-efficient fine-tuning is employed to adapt multiple mainstream LLMs. Extensive experiments demonstrate that the proposed method substantially outperforms baseline models without implicit information on implicit relation extraction tasks, achieving consistent improvements in both F1 score and recall. The advantages are particularly pronounced in high relational complexity settings and pure implicit relation scenarios, highlighting the critical role of implicit information augmentation in enhancing the reasoning capability and relational coverage of LLMs. Our code is available at https://github.com/IIP408/IMP-RELLM.

Program induction is a typical approach that helps Large Language Models (LLMs) in complex knowledge-intensive question answering over knowledge bases (KBs) to alleviate the hallucination of LLMs. However, accurate program induction requires extensive high-quality parallel data for a specific KB, which is scarce for low-resource KBs. Moreover, the heterogeneity of questions and KB schemas limits the transferability of models trained on a single dataset. To this end, we propose REPANA, a reasoning path navigated program induction framework that enables LLMs to reason over heterogeneous KBs. We decouple the program induction into perceiving the KB and mapping questions to program sketches. Accordingly, our framework consists of (1) an LLM-based navigator, which retrieves reasoning paths of the input question from the given KB; (2) and a KB-agnostic parser trained on multiple heterogeneous datasets, which takes the retrieved paths and the question as input and generates the corresponding program. Experiments show that REPANA exhibits strong generalization and transferability. It can directly perform inference on datasets not seen during training, outperforming other state-of-the-art low-resource methods, even approaching the performance of supervised methods.

Large Language Models (LLMs) demonstrate impressive natural language capabilities but often struggle with knowledge-intensive reasoning tasks. Knowledge Base Question Answering (KBQA), which leverages structured Knowledge Graphs (KGs) exemplifies this challenge due to the need for accurate multi-hop reasoning. Existing approaches typically perform sequential reasoning steps guided by predefined pipelines, restricting flexibility and causing error cascades due to isolated reasoning at each step. To address these limitations, we propose KG-Hopper, a novel Reinforcement Learning (RL) framework that empowers compact open LLMs with the ability to perform integrated multi-hop KG reasoning within a single inference round. Rather than reasoning step-by-step, we train a Reasoning LLM that embeds the entire KG traversal and decision process into a unified “thinking” stage, enabling global reasoning over cross-step dependencies and dynamic path exploration with backtracking. Experimental results on eight KG reasoning benchmarks show that KG-Hopper, based on a 7B-parameter LLM, consistently outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models such as GPT-3.5-Turbo and GPT-4o-mini, while remaining compact, open, and data-efficient.

Recent breakthroughs in large language models have made it feasible to effectively summarize cross-lingual dialogue information, proving essential for the global communication context. However, existing methodologies encounter difficulties maintaining factual consistency across multiple dialogue exchanges and lack clear explanations of the summarization process. This paper presents a novel factually consistent prompting technology with large language models to address these challenges in cross-lingual dialogue summarization. First, we propose a factual replacement mechanism to enhance information analysis by incorporating noise information into summarization candidates. We adopt a self-guidance framework to enforce factual consistency, enhancing information flow tracking in cross-lingual hybrid dialogue scenarios with the assistance of GPT-based models. Furthermore, we introduce a view-aware chain-of-thought-driven architecture to improve the interpretability and transparency of the cross-lingual dialogue summarization process. Comprehensive experimental evaluations on cross-lingual summarization tasks, spanning English, French, Spanish, Russian, Chinese, and Arabic, and hybrid cross-lingual tasks substantiate that the proposed model achieves superior performance relative to state-of-the-art baselines.

Code summarization plays a crucial role in helping developers quickly grasp the purpose and behavior of source code. However, the majority of existing approaches concentrate solely on the function level, overlooking the richer contextual signals and hierarchical relationships that naturally exist within large software repositories. To address these shortcomings, we present a new hierarchical code summarization framework built upon GraphRAG. Our key observation is that the community structures revealed by GraphRAG inherently align with the layered organization of real-world software systems—for instance, functions forming classes, and classes composing files. Leveraging this insight, we construct a multi-level code summary graph and dynamically retrieve the most relevant subgraph to supply a large language model with structured and contextually coherent information. This design alleviates the context fragmentation commonly seen in traditional retrieval-augmented generation pipelines. Comprehensive experiments demonstrate that our framework markedly improves summary quality, especially when applied to complex or large-scale codebases.

In this paper, we propose BeatMamba, a drum-separated beat tracking framework built on the Mamba state space architecture. Beat tracking in polyphonic music is challenged by complex mixtures and high computational costs. To address these issues, We first apply music source separation to isolate the drum stem, preserving salient percussive cues while suppressing interference from vocals and harmony. This representation improves rhythmic clarity and reduces ambiguity in mix audios. For sequence modeling, we employ selective state space modules to capture long-range temporal dependencies with linear-time and memory complexity, avoiding the quadratic cost of self-attention. Compared to Transformer baseline models, BeatMamba significantly improved computational efficiency and reduced model size, making it more suitable for low-latency and real-time applications. Experimental results on benchmark datasets demonstrate that BeatMamba reduces parameter counts while maintaining or even surpassing the accuracy of heavy models.

To address the challenges of non-stationary signals and the disruption of spectral continuity by standard patch-based Transformers in underwater acoustic target recognition, we propose SpecSlice-ViT. This novel architecture preserves spectral integrity via an acoustic-aware Feature Slicing strategy that slices completely along the frequency dimension to maintain harmonic structure. We introduce a dual-view fusion mechanism, combining fine-grained Short-Time Fourier Transform (STFT) details with robust Mel-Frequency Cepstral Coefficients (MFCC) envelopes through cross-attention and adaptive gating. Additionally, a Hierarchical Encoder cascades window attention with token compression to efficiently model multi-scale temporal dependencies. Experimental results on the DeepShip dataset demonstrate that SpecSlice-ViT outperforms existing Convolutional Neural Networks and Transformer baselines, achieving competitive accuracy and efficient inference speeds, indicating potential for future deployment on underwater platforms.

Automatic Speech Recognition (ASR) aims to tran scribe speech signals into textual sequences, serving as a funda mental technology in human-computer interaction. However, end to-end ASR systems often suffer from inconsistent predictions across different stochastic passes and insufficient representation learning capability. To mitigate these issues, we propose RUCL (Regularization and Unsupervised Contrastive Learning), a novel framework that hierarchically integrates consistency regulariza tion and unsupervised contrastive learning.Specifically, at the feature level, we employ unsupervised contrastive learning to enhance the discriminability of speech representations, thereby strengthening the model’s ability to capture robust acoustic features. At the output level, we apply consistency regularization to minimize the divergence between two CTC distributions generated under different dropout masks, promoting prediction stability. Extensive experiments on AISHELL-1 and LibriSpeech datasets demonstrate that RUCL achieves superior performance, with a CER of 4.63% on AISHELL-1 and WERs of 4.44%/9.96% on LibriSpeech test-clean/test-other, significantly outperforming strong baselines.

Voice Conversion (VC) modifies the speaker identity of speech while preserving its linguistic content. It has transformed the domains of voice dubbing, voice cloning, and assistive technologies by improving the naturalness of the synthesized speech. Recent advances in generative models, particularly diffusion-based frameworks, have significantly improved perceptual quality but come at a high computational cost and face challenges, including degraded prosody and reduced speech intelligibility. These primarily arise from attention-based sequence modeling, which scales quadratically. In this work, we propose SSMamba-VC, a selective state space (SS)-based hierarchical discussion framework that replaces attention mechanism with linear-time sequence modeling. Compared to the existing state-of-the-art VC systems, SSMamba-VC achieves notable improvements across \textit{subjective}, \textit{objective}, and \textit{quantitative} metrics. In particular, the proposed approach (DiffHier-VC + Ours) achieves the NISQA-MOS of 3.90 (+0.02) and a UTMOS of 2.92 (+0.16), outperforming existing diffusion baselines in perceptual stability. The proposed approach reduces $F_{0}$ RMSE to 42.44, improving prosodic accuracy, while also improving intelligibility, with WER reduced to 18.14 and CER to 8.73 compared to DiffHier-VC. Furthermore, with 98\% fewer computational resources, SSMamba-VC achieves a $\sim$15.6\% improvement in real-time factor (RTF), demonstrating that linear-time SS modeling enables faster inference without degrading the quality of the synthesized VC. SSMamba-VC further reduces the real-time factor to achieve high-fidelity VC for practical deployment. The demo utterances are publicly available at: \textit{\url{https://iamshreeji-copy2.github.io/SSMamba-VC/}}

Recent advances in diffusion and flow-matching models have enabled high-fidelity text-to-image synthesis. However, conditioning these models on audio remains challenging, as token-level fusion through cross-attention increases the attention length and introduces substantial computational overhead. We present AudioControl, an efficient audio-to-image conditioning module for flow-matching generators that adapts global affine conditioning for audio guidance with minimal overhead. AudioControl extracts a compact global audio embedding with a pre-trained Audio Spectrogram Transformer (AST) and maps it through a lightweight projection to obtain modulation vectors. These vectors are fused additively with pooled text embeddings to form a unified global condition, which is then injected into Transformer blocks via adaptive layer normalization. By modulating feature statistics instead of concatenating audio tokens, AudioControl preserves the original attention sequence length and avoids the quadratic scaling associated with longer contexts, while still steering the full visual latent space. Experiments on VGGSound and GreatestHits demonstrate improved audio-visual alignment with competitive image quality, while substantially reducing the parameter footprint and inference latency relative to token-based conditioning strategies.

Automatic singing assessment relies on aligning sung notes with a reference score; however, achieving robust alignment remains challenging for beginners due to pitch instability and phrasing errors. We propose AnchorAlign++, an end-to-end framework that explicitly models the number of notes aligned to each lyric token as a structural variable and converts this dis- crete prediction into actionable temporal constraints through an anchor-driven inference strategy. The model consists of a shared acoustic encoder, a non-autoregressive note decoder for frame- level pitch and onset estimation, and an autoregressive lyric decoder that jointly predicts lyric tokens and their corresponding note counts. To further enhance robustness, we introduce a unidi- rectional acoustic–symbol fusion mechanism that injects frame- level pitch and onset cues into lyric decoding, thereby stabilizing long-tail note count prediction. We further propose a count– onset consistency loss, which couples discrete note counts with continuous onset probabilities to reduce boundary drift. During inference, the predicted note counts guide the selection of onset anchors, forming sharp lyric–note boundaries without manual correction. Experiments on multiple singing datasets demonstrate that AnchorAlign++ consistently improves transcription accuracy and alignment robustness, while maintaining practical compu- tational efficiency for deployment, providing reliable automatic singing assessment for both novice and professional singers.

Despite recent progress in generating natural and personalized speech, voice conversion (VC) systems that aim to modify paralinguistic features of speech while preserving linguistic content remain vulnerable to degraded inputs, where even minor adversarial perturbations cause severe performance drops. Prior studies mainly focused on perturbations that affect speaker identity in classic encoder-decoder models (e.g., AdaIN-VC, AutoVC), while the impact of linguistic-targeted perturbations, especially on modern neural audio codecs, remains unexplored. To address this gap, we propose LM-AdvVC, a two-stage localized adversarial attack framework that applies speaker-guided perturbations directly to content embeddings. Firstly, a critical-segment local ization with phoneme- and word-level analysis adaptively identifies and perturbs vulnerable regions. Secondly, the waveform perturbations are refined using CMA-ES optimization under a teacher forcing cross-entropy objective, yielding transferable adversarial attacks across gray-box scenarios. Experimental results demonstrate that LM-AdvVC successfully manipulates linguistic content under both white and gray-box settings and reveal the content vulnerability of VC systems. The codes are available at https://github.com/windforestfiremountain/LM-AdvVC.

The Audio-Visual Question Answering (AVQA) task aims to mine temporal dynamic information in videos and perform comprehensive spatio-temporal reasoning to answer given questions accurately. Although pre-trained models have achieved significant success in audio-visual question-answering tasks, they still face several challenges. On the one hand, the length of long-term videos makes it difficult for image-based and audio-based pre-trained models to capture temporal dynamic information effectively. On the other hand, the textual semantic gap between the question and the original text of the text-based pre-trained models limits the prediction results. To address the above challenges, we propose a Temporal Dynamics Perception Model (TDPM) to capture temporal dynamic dependencies in videos through a visual temporal alignment module and an audio temporal alignment module. Those temporal alignment modules are constructed via a language-guided auto-regressive task and can combine historical clues of videos and language guidance to predict the future state of videos. Also, TDPM mitigates the semantic gap between the original text of the question and the pre-trained models through a textual semantic alignment module. Experimental results on the MUSIC-AVQA dataset demonstrate the effectiveness of our proposed model. It outperforms all the compared models in terms of average accuracy, with a maximum improvement of 17.02% and an average improvement of 10.86%.

In this paper, we propose a novel approach for solving the interval data-based multi-objective non-linear fixed-cost transportation problem with different means of transport by integrating priority structures into the framework of the NSGA-II. We analyzed different ways of calculating the crowding distance corresponding to a solution of the investigated problem and proposed new genetic operators specifically designed to handle different means of transport. The computational experiments on the existing instances from the literature show that our approach is suitable and efficient for solving the considered multi-objective non-linear fixed-cost transportation problem.

Expensive optimization problems with mixed contin-uous and categorical variables (EOPMVs) arise in engineering and scientific design, yet remain difficult because evaluations are costly and the two variable types interact in non-smooth ways. To address this challenge, we propose multiform transfer ant colony optimization (MFT-ACO), a surrogate-assisted framework grounded in the multiform transfer optimization paradigm. MFT-ACO reformulates the target problem into three complementary auxiliary formulations by exploiting the distinct modeling strengths of different surrogates for heterogeneous variables: an RBF-based full-space formulation for smooth interpolation, an LSBT-based combinatorial formulation for categorical discontinuities, and a category-conditioned local refinement formulation for continuous improvement in promising regions. These forms co-evolve through an iterative multiform selection mechanism that enables additional surrogate-guided refinement before expensive evaluations are performed. Experimental results on benchmark EOPMVs show that MFT-ACO is competitive under a limited evaluation budget, suggesting that multiform reformulation is a viable strategy for expensive mixed-variable optimization.

Evolutionary algorithms have demonstrated strong performance on combinatorial problems such as Sudoku; however, their effectiveness heavily depends on handcrafted local search operators and fixed control heuristics. Designing adaptive operator control mechanisms that generalize across problem instances remains a challenging and largely manual process. This paper proposes an LLM-evolved adaptive operator control framework for Sudoku solving, in which a large language model (LLM) is integrated into the evolutionary loop to evolve state-aware local search control heuristics. Rather than directly constructing solutions or performing search, the LLM generates symbolic heuristic policies that regulate swap selection and mutation pressure based on runtime indicators of the search process. These heuristics are encoded as lightweight control rules and embedded into a local search-based genetic algorithm (LSGA). To ensure robustness and mitigate hallucination effects, all LLM-generated heuristics are subjected to an empirical verification procedure across multiple Sudoku instances, and only validated heuristics are retained through an evolutionary selection mechanism. Experimental results on standard Sudoku benchmarks demonstrate that the proposed method achieves competitive performance compared to state-of-the-art approaches and consistently improves convergence speed relative to the original LSGA.

The node combinatorial optimization (NCO) problem in complex networks tries to select a set of critical nodes to achieve objectives such as influence maximization, robustness enhancement, and edge coverage, etc. Owing to its significant applications in social network analysis, infrastructure protection, and system design, NCO has attracted extensive research attention. Recently, evolutionary deep reinforcement learning (EDRL) emerges as a promising approach by reformulating discrete NCO into continuous parameter optimization of deep Q networks. However, existing EDRL studies primarily focus on single-task scenarios, while how to efficiently solve multiple NCO problems simultaneously remains a significant challenge. To address this issue, we introduce a multi-gate mixture-of-experts (MMoE)-based multifactorial EDRL algorithm (called MMoE-EDRL) that employs a MMoE-based deep Q network (called MDQN) to model the multitask NCO problem as a weight optimization task. Specifically, the MDQN employs shared expert layers to learn general node representations across tasks and equips each task with an independent gating network to explicitly capture task similarities and differences. Subsequently, this MDQN is evolved by integrating EDRL with the multifactorial evolutionary algorithm (MFEA) for effective multitask learning (MTL). Experimental results across six real-world networks demonstrate the efficiency of MMoE-EDRL over state-of-the-art methods in addressing the NCO problems. This work provides an efficient solution for multitask node combinatorial optimization in complex networks.

For time series databases, single-column storage and single-column-group storage are common storage schemes. The former stores each time series independently but results in redundant storage of timestamps. The latter stores a batch of time series by sharing a time column but requires extra disk space for recording null values. This paper aims to design an algorithm that finds a compromise storage scheme between the two. We divide a batch of time series into multiple column-groups to minimize the overall space cost. We first model the time series grouping problem as an optimal set partitioning problem. Subsequently, we propose a two-phase algorithm to solve this problem. In Phase 1, we propose a similarity metric based on the time overlap degree between time series, and we employ a clustering algorithm to generate initial solutions. In Phase 2, based on the genetic algorithm (GA), we propose genetic operators derived from the features of time series to improve the quality of solutions obtained in Phase 1. Finally, we validate the effectiveness of the proposed algorithm through experiments.

Open-pit mine scheduling is a complex real-world optimization problem that involves uncertain economic values and dynamically changing resource capacities. Evolutionary algorithms are particularly effective, as they can easily adapt to such environments. However, uncertainty and dynamic changes are often studied in isolation in real-world problems. In this paper, we study a dynamic chance-constrained open-pit mine scheduling problem where block economic values are stochastic and mining and processing capacities vary over time. We adopt a bi-objective evolutionary formulation that simultaneously maximizes expected discounted profit and minimizes its standard deviation. To address dynamic changes, we propose a diversity-based change response mechanism that repairs a subset of infeasible solutions and introduces additional feasible solutions whenever a change is detected. We evaluate the effectiveness of this mechanism across four multi-objective evolutionary algorithms and compare it with a baseline re-evaluation–based change-response strategy. Experimental results demonstrate that the proposed approach consistently outperforms the baseline methods across different uncertainty levels and change frequencies.

Constrained multi-objective optimization (CMOP) is challenging because the search must balance feasibility maintenance, convergence toward the constrained Pareto-optimal set, and diversity preservation under complex constraints and limited evaluation budgets. This paper proposes Closed-Loop Success-History Driven Reconstructed Differential Evolution(CLSHRDE), which is built upon the Reconstructed Differential Evolution (RDE) baseline under the Institute of Electrical and Electronics Engineers (IEEE) Congress on Evolutionary Computation (CEC) 2025 CMOP benchmark setting, where RDE has demonstrated strong overall performance. On this basis, CLSHRDE introduces an explicit two-operator search structure by integrating an extra stochastic mutation rule with the original order-based mutation. To enable robust stage wise resource allocation between the two operators, CLSHRDE develops a closed-loop scheduling mechanism driven by an exponentially weighted success-history estimate, providing noise attenuated feedback for online operator regulation across search stages. Extensive experiments on the Scalable Decision space Constraints (SDC) benchmark suite and real-world constrained multi-objective problems (RWMOPs) show that, under unified parameter settings, CLSHRDE achieves competitive or improved performance in terms of Inverted Generational Distance (IGD), Hypervolume (HV), and feasible success rate (FSR), with improvements statistically supported by one-sided Wilcoxon rank sum tests.

Hate speech detection faces significant challenges due to inconsistent criteria across datasets, rooted in inherent subjectivity and annotation bias. This paper argues that the core of this task lies in deep adaptation to specific annotator stances within particular scenarios rather than merely fitting data distributions across dataset. Inspired by the Dual-Process Theory in human cognitive science, we propose DCAdapt, a dual-cognition driven framework for adaptive hate speech detection that achieves high adaptability by simulating the synergy between intuitive association and logical calibration. The framework decomposes the detection task into two collaborative stage: decoupled dual-knowledge base construction and dual-pathway cognitive reasoning. In the construction stage, a mutual-indexing mechanism decouples objective knowledge from subjective precedents to achieve synergistic modeling of general knowledge and annotation criteria. In the reasoning stage, the framework emulates the decision sequence from affective association to rational judgment: an implicit pathway leverages a general knowledge base to uncover latent contextual clues, while an explicit pathway employs a specialized knowledge base to enforce semantic constraints and stance alignment based on precedents for informed adjudication. This dual-pathway coordination enables DCAdapt to guide Large Language Models (LLMs) to precisely identify complex decision boundaries and align with target-scenario annotation stances. Experimental results demonstrate that DCAdapt exhibits superior effectiveness, architectural flexibility, and cross-scenario transferability, particularly in detecting complex and implicit hate speech.

Explainable recommendation systems aim to provide both recommendation results and corresponding explanations to enhance user trust and engagement. Existing post-hoc methods leverage pre-trained language models to generate textual explanations. However, these works hardly consider the latent topical structures in user reviews, resulting in generic and repetitive explanations with limited personalization. To address this issue, this paper introduces TPLER (Topic-enhanced Prompt Learning for Explainable Recommendation), a novel model that integrates topical preferences into prompt learning to generate more personalized and higher-quality explanations. Our approach employs a topic-enhanced feature extraction strategy based on Latent Dirichlet Allocation (LDA) to explicitly model user and item topical preferences from review texts. These topic-aware features are fed into a pre-trained transformer model via continuous and discrete prompts, guiding the generation process. Continuous prompts encode rating information to capture personalized user-item interactions, while discrete prompts are obtained from topic-enhanced features. A two-stage multi-task training strategy is proposed to jointly optimize the rating prediction and explanation generation tasks. Experiments on three benchmark datasets show that TPLER achieves an average improvement of 7.93% across text quality metrics (BLEU-1/4, ROUGE-1/2/L) and 21.63% for personalization metrics, verifying the effectiveness of our proposed TPLER model.

Deep learning--based static malware detection models face severe supply chain security threats. Existing defense strategies against backdoored static malware detection models are largely built upon a core assumption: that fine-tuning contaminated models or retraining them on clean data can effectively remove backdoors. However, our study reveals critical limitations of this widely trusted security assumption. To expose this defense blind spot, we are the first to propose a Model Architecture Backdoor attack in the domain of static malware detection.Considering the strict format and semantic constraints of binary files, we design a universal architecture-level backdoor scheme based on a specific trigger mechanism. Unlike traditional data poisoning or weight manipulation attacks, the proposed attack introduces malicious structural components that are independent of model weights, enabling the backdoor to survive and persist even after full retraining in a clean environment. Extensive experiments conducted on the SOREL-20M and MOFIT benchmarks demonstrate that mainstream architectures, including CNNs, LSTMs, and Transformers, remain vulnerable to this architecture-level attack despite rigorous retraining procedures. Our findings expose the inherent limitations of defenses that rely solely on weight verification and highlight the necessity of incorporating architecture auditing and structure-aware defenses into the model supply chain to enhance the security and trustworthiness of AI systems in real-world deployment scenarios.

The rapid development of Large Language Models (LLMs) has raised concerns about the authenticity of LLM-generated text, and existing multi-bit watermarking schemes have low robustness to cropping, limited capacity, and high false positives. We propose a robust framework based on clustered multi-bucket coding. First, for token embedding, we use frequency-balanced adaptive k-means for unsupervised clustering, which achieves the goal of improving the robustness of the framework. Then, in each cluster, we use multiple buckets to subdivide again to form a combination of red and green buckets, which in turn improves the embedding capacity. Finally, we employ two-stage detection with statistical validation and combinatorial decoding to achieve a low false positive rate. Experiments on GPT-2 and OPT-1.3B show that our framework outperforms state-of-the-art methods in terms of robustness ($92.65\%$ accuracy after $50\%$ cropping), capacity ($96.33\%$ at 20 bits), and false positives ($\le 10^{-6}$), providing an effective solution for traceable LLM-generated text.

Large Language Model (LLM) agents are increasingly integrated into critical systems, leveraging external tools to interact with the real world. However, this capability exposes them to Indirect Prompt Injection (IPI), where attackers embed malicious instructions into retrieved content to manipulate the agent into executing unauthorized or unintended actions. Existing defenses predominantly focus on the pre-processing stage, neglecting the monitoring of the model's actual behavior. In this paper, we propose PlanGuard, a training-free defense framework based on the principle of Context Isolation. Unlike prior methods, PlanGuard introduces an isolated Planner that generates a reference set of valid actions derived solely from user instructions. In addition, we design a Hierarchical Verification Mechanism that first enforces strict hard constraints to block unauthorized tool invocations, and subsequently employs an Intent Verifier to validate whether parameter deviations are benign formatting variances or malicious hijacking. Experiments on the InjecAgent benchmark demonstrate that PlanGuard effectively neutralizes these attacks, reducing the Attack Success Rate (ASR) from 72.8% to 0%, while maintaining an acceptable False Positive Rate of 1.49%. Furthermore, our method is model-agnostic and highly compatible.

The increasing integration of artificial intelligence (AI), and particularly machine learning (ML), into software-intensive systems raises security concerns for AI/ML-specific code. Although AI-based vulnerability detection models (VDMs) achieve accurate prediction on contemporary software, their cross-domain effectiveness on AI/ML-specific software remains underexplored due to limited labeled data. To fill this gap, in this paper we aim to examine whether VDMs trained on contemporary software can accurately detect vulnerabilities in AI/ML-specific software as well, and to assess how domain adaptation can affect their performance. We construct a dataset of vulnerable and non-vulnerable Python functions split into contemporary and AI/ML-specific subsets, fine-tune CodeBERT and CodeGPT on the contemporary subset, and evaluate them on the AI/ML subset. We then perform gradual adaptation by adding 5%, 10%, 15%, 20%, 25%, 40%, and 50% target- domain samples. We also analyze representation drift with UMAP and attention drift via self-attention-based explainability. Results suggest that AI-based VDMs built on contemporary software cannot be used effectively on AI/ML-specific software without proper domain adaptation, indicating that AI/ML-specific soft- ware exhibits unique characteristics; however, even a small- moderate amount of domain-specific data can enable satisfactory cross-domain detection accuracy

Adversarial attacks pose serious security risks to deep neural networks (DNNs) by inducing erroneous predictions. While gradient-based and GAN-based attack methods have achieved high success rates, they often produce perturbations that are diffusely distributed across images, resulting in poor imperceptibility. This limitation restricts their applicability in scenarios such as privacy protection for face recognition systems. To address this issue, we propose SpAdvGAN, a semantic-aware adversarial example generation framework that integrates sparse attention with generative adversarial networks. SpAdvGAN introduces a sparse semantic attention extractor that explicitly constrains perturbations to a limited set of semantically important regions, preventing unnecessary noise diffusion and preserving visual fidelity. This design enables precise regulation of perturbation granularity while maintaining strong attack effectiveness. Extensive experiments on CIFAR-10 and ImageNet datasets demonstrate that SpAdvGAN substantially improves the imperceptibility of adversarial examples, while achieving comparable or higher attack success rates than state-of-the-art gradient-based attacks and the GAN-based baseline LpAdvGAN.

Logs serve as the foundational audit trails for monitoring the integrity and security of modern software systems. Automated log anomaly detection is crucial for ensuring the reliability of modern software infrastructures. However, existing approaches encounter significant limitations. Traditional deep learning models frequently lack interpretability and struggle with generalization, whereas Large Language Models are prone to hallucinations, introducing new security risks through factual fabrication. To address these challenges, we propose RationAnomaly, a novel framework designed to enhance log anomaly detection by synergizing Chain-of-Thought fine-tuning with reinforcement learning. Our approach initially instills expert-level reasoning patterns using supervised fine-tuning guided by Chain-of-Thought data, grounded in a high-quality dataset corrected through a rigorous expert-driven review process. Subsequently, we introduce a reinforcement learning phase employing a multi-faceted reward function to optimize the model. This function simultaneously evaluates format adherence, answer accuracy, and logical consistency, thereby effectively mitigating hallucinations. Experimental results demonstrate that RationAnomaly outperforms state-of-the-art baselines by achieving superior F1-scores on key benchmarks while providing transparent and step-by-step analytical outputs.

Smart contracts are immutable once deployed, making vulnerabilities prone to severe and irreversible losses. Existing deep-learning--based detectors often operate as black boxes and fail to capture vulnerability-relevant semantics underlying real attack mechanisms. We propose MRV-GCN (Multi-Relational Vulnerability-aware Graph Convolutional Network), a graph-based framework that integrates explicit expert knowledge into deep representation learning for smart contracts. MRV-GCN constructs a multi-relational semantic graph (MRSG) that explicitly models AST structures together with control- and data-flow dependencies, and introduces hierarchical vulnerability semantic features (HVSF) to inject expert-defined cues---such as external calls, state updates, unsafe arithmetic, and delegatecall chains---directly into node representations. An interpretable hybrid readout (R-Hybrid) module further fuses semantic subgraph evidence with non-semantic context to produce robust predictions. Experiments on real-world datasets demonstrate that MRV-GCN achieves strong performance across multiple vulnerability types, while the learned fusion coefficient α provides a lightweight semantics-aware signal for interpretation by quantifying the influence of expert-defined semantic evidence in each prediction.

Heterogeneous Graph Neural Networks (HGNNs) have become essential for encrypted traffic analysis, yet existing backdoor attacks fail to exploit their vulnerabilities due to a fundamental mismatch with homogeneous assumptions. Current methods rely on additive assumptions that fail to capture cooperative interactions for precise localization, suffer from signal dilution caused by massive structural redundancy, and generate semantically invalid triggers that violate rigid network protocols.

To bridge this gap, we propose \textbf{MEHGB} (\textbf{\emph{M}}eta-Path Pruned and \textbf{\emph{E}}xplainable \textbf{\emph{H}}eterogeneous \textbf{\emph{G}}raph \textbf{\emph{B}}ackdoor), a lightweight framework designed to expose the vulnerabilities of NIDS. The method uniquely integrates meta-path guided pruning to distill the informative graph core, a Conditional Random Field (CRF) explainer to capture non-linear cooperative interactions for precise localization, and constraint-driven optimization to enforce strict structural and statistical compliance. Empirical evaluations on UNSW-NB15, CSE-CIC-IDS2018, and CTU-Mixed-Capture demonstrate that MEHGB outperforms state-of-the-art baselines by 9.6\% in Attack Success Rate while reducing trigger generation latency by 53.6\%. Furthermore, the framework exhibits strong task-agnostic transferability and resilience against advanced defenses. Ultimately, by exposing these latent vulnerabilities, the study serves as a critical stress test to guide the development of next-generation NIDS architectures resilient to complex heterogeneous threats.

Traditional Chinese Medicine (TCM) herbal prescription recommendation involves selecting appropriate herb combinations from patient symptoms under symbolic constraints derived from TCM theory. While most data-driven approaches achieve strong predictive accuracy, they typically encode domain knowledge only implicitly and lack explicit mechanisms to verify and correct knowledge-inconsistent predictions, which limits their reliability for clinical decision support. In this paper, we propose SEGCN-AIKC, a Semantic-Enhanced Graph Convolutional Network with Adaptive Inconsistency-aware Knowledge Correction, which formulates TCM herbal recommendation as a constrained hypothesis inference problem. Specifically, a semantic-enhanced multi-view graph convolutional network captures heterogeneous symptom–herb relations and constructs a syndrome-level hypothesis, which is then verified and minimally corrected by AIKC under symbolic TCM constraints. By integrating neural perception with constraint-guided verification and repair, SEGCN-AIKC bridges data-driven learning with symbolic reasoning, yielding recommendations that are both accurate and theoretically consistent. Experiments on a public TCM prescription dataset demonstrate consistent improvements over baselines, while ablation and case studies further validate interpretability and clinical plausibility.

Graph Anomaly Detection (GAD) aims to identify anomalous nodes whose behavioral patterns or structural relationships deviate from the majority, a task that has attracted significant research attention. However, constructing a universal GAD framework capable of adapting to diverse real-world applications faces the following technical bottlenecks: on one hand, existing models struggle to achieve a balance between capturing complex long-range dependencies and maintaining high computational efficiency; on the other hand, mainstream methods are constrained by "generalization silos," lacking the cross-domain capability to handle data distribution shifts. To alleviate these issues, we propose AceGAD, a universal graph anomaly detection method. Specifically, AceGAD consists of two core components: first, to address the dilemma of reconciling long-range dependencies with computational efficiency, we designed the Linear Graph State-Space Encoder (LG-SSE), which reformulates graph propagation as a continuous linear state evolution process, achieving precise memory of global context while maintaining linear time complexity; second, to break through the limitations of generalization silos, we proposed the Context-Aware Metric Projector (CAM-Pro), which enables the model to adapt to new environments using only a few normal samples by constructing an adaptive metric space, achieving Zero-Fine-Tuning inference on new graphs. Extensive experiments conducted on eight public benchmark datasets across different domains demonstrate that AceGAD achieves superior performance, efficiency, and universality.

Cross-domain image segmentation plays a crucial role in the field of image segmentation, as it helps to improve the generalization ability of models across different datasets. However, existing models overemphasize domain-invariant features that are not specific to a particular domain, such as cells in medical images or buildings in remote sensing images. Consequently, they neglect the importance of domain-specific features, such as object contours or boundaries.Therefore, we propose Balanced Domain SAM(BDSAM), a model designed to balance domain-specific features while enhancing domain-invariant features.Specifically, we introduce an Iterative Optimization Selection Module (IOSM). By leveraging query features to continuously filter for more relevant pixels within the support features, we automatically select these pixels as point prompts. Furthermore, we propose a Domain-specific Feature Fusion Module (DFFM). By extracting and fusing features from different levels, we enhance the focus on domain-specific features.Extensive experiments on four cross-domain datasets show that our model has better average accuracy on 1-Shot and 5-Shot settings than the state-of-the-art approaches respectively .

Session-based recommendation (SBR) aims to predict the next item based on anonymous user sessions. However, existing studies still face challenges below: \textbf{(1)} Inadequate capture of both pairwise and high-order relationships in graph relations. \textbf{(2)} Spurious edges in the global graph lead to redundant global relations extraction. \textbf{(3)} Complex coupling of different user intents caused by noise and dynamics within sessions. To address these issues, we propose a novel \underline{H}igher-order \underline{H}ybrid Graph Convolutional Network (H$^2$-GCN) for SBR, which effectively captures multi-level relationships and refines diverse user intents, enabling effective recommendations. Specifically, we first present a higher-order graph based on simplicial complexes (SCs) to learn the multilevel complex dependencies of items and sessions separately. Moreover, two intent expert modules are designed to independently learn user intents within sessions: the short-term intent expert module is proposed to block long-term dependencies and activate true short-term intents from user noise behaviors, while the long-term intent expert module employs a dynamic-static graph collaborative propagation to filter redundant global relations and derive true long-term intents. Eventually, we guide the intents obtained from the two expert modules to form the final user intents. Extensive experiments show that H$^2$-GCN outperforms state-of-the-art methods in three real-world datasets.

Log anomaly detection is essential to the reliability of large-scale software systems. In real-world deployment, anomaly alarms should not only be accurate but also reveal when anomalies occur, where they arise, and which components and events are most responsible. However, most existing methods model logs from a single view and lose fine-grained traceability during representation learning, which limits their value for actionable diagnosis. To address this issue, we propose IP-Log, an attribution-aware framework for log anomaly detection based on index-preserving multi-view fusion. IP-Log models system behavior from three complementary views, namely component, event, and time, and structures their interactions according to two principles. One is level-confined information flow, which prevents premature entanglement across views. The other is index-preserving computation, which maintains the correspondence between latent representations and original log entities throughout the model. As a result, anomaly evidence can be decomposed into fine-grained contributions over components, events, and time segments, and the model can further identify top-ranked evidence triplets for diagnosis and troubleshooting. Experiments on public log benchmarks show that IP-Log achieves strong anomaly detection performance while also providing interpretable and index-aligned evidence attribution.

Social recommendation plays a critical role in modern recommender systems. Graph-based methods effectively model relational signals to improve performance. However, real-world interaction data contains abundant noise, while social links remain typically sparse and unreliable. These limitations significantly hinder the robustness of current models. Existing research rarely explores implicit social relations induced by user preferences. To address these challenges, Prior and Implicit Reliability Graph Learning is proposed. This unified framework explicitly models reliability across interaction, relation, and fusion levels. Specifically, the model integrates a self-paced interaction denoising strategy with a hybrid social graph containing preference-based neighbors. A reliability-aware fusion gate further balances these heterogeneous user representations. Experiments on three real-world benchmarks demonstrate consistent improvements in Recall and NDCG. These findings establish reliability-oriented graph learning as a principled direction for advancing robust social recommendation.

Ensuring both safety and efficiency in decision-making of autonomous driving systems remains a fundamental challenge. Traditional Deep Reinforcement Learning (DRL) suffers from unsafe random exploration and slow convergence, while Large Language Models (LLMs) demonstrate inherent latency in real-time inference operations. To address these limitations, this paper proposes SARAD, a novel safety-aware hybrid framework that synergizes LLMs and DRL for autonomous driving. SARAD substitutes DRL's random exploration with Retrieval-Augmented Generation (RAG)-enhanced, LLM-guided decisions sourced from a dynamic expert knowledge repository. An attention discriminator that integrates LLM’s prior knowledge into DRL policy optimization is proposed. A collision predictor module, which is fine-tuned with historical collision data, is further designed to guarantee the safety of the vehicle. Extensive experiments show that SARAD achieves significant performance improvements in Highway-Env simulator, validating the effectiveness of the model in autonomous driving.

Under adverse environmental conditions such as fog, rain, snow, and low illumination, object detection models often suffer significant performance degradation in real-world deployments due to domain shifts and visual feature degradation. To address this problem, this paper adopts YOLOv8 as the baseline detector and proposes a robust framework, termed DWYOLO, which is designed to learn domain-invariant object representations in complex environments.At the core of the framework is the proposed Dual-Context Attention Module (DCAM). This module integrates short-distance attention and long-distance attention with multi-scale contextual fusion, enabling simultaneous modeling of global contextual relationships and local fine-grained details. By explicitly leveraging complementary contextual information, DCAM substantially improves the robustness of the characteristics against weather-induced domain drift. Furthermore, an adaptive optimization strategy, Weighted Intersection over Union (WIoU), is incorporated to emphasize challenging and underrepresented samples during training. This strategy improves generalization across diverse environmental conditions while preserving an end-to-end training pipeline and without incurring significant computational overhead. Extensive experiments on multiple public benchmark datasets demonstrate that DW-YOLO consistently outperforms baseline detectors under adverse weather conditions, yielding notable improvements in both detection accuracy and robustness. These findings indicate that DW-YOLO provides an effective solution to object detection under domain shift and shows strong potential for deployment in real-world intelligent perception systems.

Raw point cloud data collected in real-world scenarios often encounter complex interferences such as sensor noise and uneven density, making high-fidelity point cloud completion tasks extremely challenging. Current mainstream point cloud completion methods generally suffer from insufficient noise resistance, struggling to meet practical application requirements. Moreover, publicly available datasets for noisy point cloud completion are relatively scarce. We proposes a point cloud completion network via the latent space-driven two-stage noise synthesis and restoration strategy, constructing a novel unified framework. Within this framework, noise synthesis and restoration are achieved through gradient guidance and constrained projection mechanisms under the constraint of hyperspherical distribution in the feature-level latent space. Experimental results on public datasets demonstrate that our method significantly improves both noise robustness and completion accuracy compared to the current state-of-the-art (SOTA) techniques, offering a new solution for noisy point cloud completion tasks.

The semantic understanding of occlusion relationships is essential for applications such as autonomous driving, robot navigation, and object grasping, yet existing scene understanding methods often lack this capability. Recognizing this limitation, this paper proposes a unified framework of occlusion reasoning and de-occlusion for open scene understanding. This framework consists of three modules: 1) Visible-region-based scene understanding: the framework leverages RAM-BLIP-Grounded-SAM to automatically generate segmentation masks and semantic information for the visible regions of objects. 2) Occlusion reasoning: an improved partial completion network (PCNet) with an occlusion completion prior module is proposed to ensure robust zero-shot generalization in open scenes through a self-supervised manner, and achieves occlusion ordering with semantic information. 3) Automatic de-occlusion: analyzing occlusion ordering results, the framework enables automatically amodal segmentation and completion based on pix2gestalt, eliminating the need for manual prompts. The results on COCOA dataset show that improved PCNet achieves an occlusion ordering accuracy of 91.534%, representing a 4.357% improvement. Moreover, the unified framework for occlusion reasoning and de-occlusion demonstrates outstanding performance in open scenes.

variations, and non-stationary motion caused by platform jitter and detection noise. Most online tracking-by-detection pipelines mainly rely on frame-wise matching with limited temporal evidence, where unreliable motion priors and noisy observations can easily trigger identity drift and trajectory fragmentation. To tackle these issues, we propose AeroGraph, a window-level largegraph association framework for online UAV MOT. Specifically, we formulate data association as inference on a unified largegraph within a sliding window, which jointly models intra-frame structural context and inter-frame correspondences to aggregate multi-frame cues. Moreover, an uncertainty-guided motion fusion strategy is designed to adaptively balance motion consistency and graph affinity according to motion reliability, preventing error propagation under jitter and detection noise. In addition, a trajectory memory with window-level consistency learning is introduced to stabilize long-term identity representations and mitigate feature drift during online updates. Extensive experiments on representative UAV tracking benchmarks demonstrate that AeroGraph consistently improves association robustness and achieves competitive performance under challenging aerial conditions.

Trajectory prediction is a critical component for autonomous vehicles (AVs) to execute safe planning. While extensive work has been done on adversarial attacks against deep learning-based trajectory prediction models, most approaches generate adversarial trajectories by adding noise to trajectory points. This approach often results in unsmooth, or even jagged, trajectories, which suffer from poor stealthiness and are impractical for real-world implementation. To address this limitation, we propose a novel adversarial attack method based on Bézier Curves, named BC-Attack (Bézier Curves Attack). BC-Attack constructs adversarial trajectories by searching for control point positions. It leverages the continuous curvature of Bézier curves to ensure the generated adversarial trajectory naturally possesses geometric smoothness, while applying both soft and hard constraints to limit the offset error, velocity, and acceleration of the adversarial trajectory, improving its feasibility in the real world. Experiments on the Apolloscape and nuScenes datasets show that BC-Attack improves the trajectory prediction error by an average of 210\% compared to normal prediction. Visualization of the trajectories also demonstrates that the generated adversarial trajectories are more stealthy and smoother than other attack methods.

Accurate traffic flow forecasting is a cornerstone of Intelligent Transportation Systems (ITS).However, existing methods struggle to strike a balance between performance and efficiency: high-performance models (e.g., Transformers) achieve superior accuracy yet suffer from quadratic computational complexity O(N^2), limiting their scalability for real-time deployment; conversely, efficient lightweight models often neglect high-frequency details or global spatial information, consequently compromising prediction accuracy. To address this dilemma, we propose ProSpectNet, a novel framework that integrates spectral analysis with prototype learning. Specifically, we introduce: (1) A Multi-Frequency Decomposition(MFD) module, which explicitly disentangles traffic series into trend and detail components via gated fusion to prevent frequency aliasing; and (2) A Latent Prototype Router (LPR), which captures global semantic dependencies by routing information through learnable prototypes with linear complexity (O(NK)). Extensive experiments on four real-world traffic datasets (PEMS03, PEMS04, PEMS07, and PEMS08) demonstrate that ProSpectNet consistently outperforms state-of-the-art baselines in terms of long-term accuracy and efficiency.It achieves competitive prediction accuracy comparable to state-of-the-art Transformers (e.g., STAEformer) while outper-forming recent efficient models (e.g., STG-Mamba). Crucially,ProSpectNet reduces inference latency by approximately 14× and memory footprint by 88% compared to high-capacity baseline models, making it an ideal solution for large-scale, real-time ITS applications, which serves as a critical macro-level prior for downstream multi-agent autonomous driving and generative AI-based traffic simulations.

The accurate and efficient Vectorized high-definition (HD) Map perception is a fundamental prerequisite for reliable planning and decision-making in autonomous driving. To enhance the perception of complex road geometries and temporal dynamics, dominant paradigms either leverage explicit map priors or employ sequential modeling. Nevertheless, these two methods suffer from inherent weaknesses: the former is highly sensitive to prior quality, while the latter is prone to error accumulation, both of which ultimately limit the model's accuracy and robustness. To this end, we propose HMF-MapNet, a robust framework featuring a parallel dual-path architecture that achieves an optimal trade-off between efficiency and performance. Specifically, the geometric path utilizes a historical map bank that interfaces directly with the BEV feature map to bolster its geometric reasoning. Meanwhile, the temporal path employs a lightweight recurrent mechanism that iteratively decodes static map data over time, robustly separating static and dynamic elements.HMF-MapNet not only significantly reduces memory usage but also achieves a remarkable +6.3 mAP gain over SOTA methods under extreme rotation noise , proving its reliability for real-world deployment. Extensive experiments on nuScenes and Argoverse2 show that our method is remarkably superior to other existing methods.

Automatic sleep staging is essential for diagnosing and treatment of sleep disorders. However, most convolutional based methods rely on empirically designed architectures and overlook the intrinsic time-frequency structure of EEG signals, resulting in elevated computational complexity, limited generalization and insufficient clinical interpretability that restrict both edge deployment and clinical adoption. Addressing these challenges, we propose an explainable and lightweight network LKSleepNet that integrates physiological priors into its architecture. The Fourier Transform Convolution (FTConv) and Large Kernel Convolution (LKConv) jointly capture global time-frequency patterns by combining transfer-domain filtering with large receptive field modeling. The lightweight 2D temporal folding module (LKTimes) provides efficient contextual representation by folding temporal information into a compact 2D feature map. The Prior-Guided Spectral Attention (PGSA) module actively extracts stage-discriminative spectral patterns by generating learnable attention masks. The spectral constraint loss regularizes feature learning to ensure physiological consistency with AASM guidelines, effectively supervising the model to focus on valid frequency bands. Furthermore, we employ Grad-CAM and Conditional Random Fields (CRF) to visualize backbone features, revealing the contributions of different frequency bands to each sleep stage. Experiments on three public datasets Sleep-EDF-20, Sleep-EDF-78, and SHHS demonstrates that LKSleepNet outperforms state-of-the-art methods, achieving macro F1-scores of 78.5%, 75.6%, and 79.5%, and overall accuracies of 85.8%, 83.2%, and 86.9%, respectively. Experimental results demonstrate the capable of LKSleepNet for robust and interpretable sleep staging on edge devices.

Single-cell annotation plays a critical role in developmental biology, disease research, and precision medicine. However, most existing approaches rely on large-scale labeled datasets for supervised training and exhibit limited generalization in few-shot settings due to the scarcity of rare cell samples, sequencing-induced technical noise, pronounced intra-class representation bias, and weak inter-class separability. To address these challenges, we propose COMET, a few-shot single-cell annotation method based on class-consistent single-cell quantization. By introducing class-consistency constraints into the quantization process, COMET maintains mutually decoupled and structured latent code domains for each cell type, discretizes continuous single-cell representations in latent space, and constructs high-purity metacell prototypes through representation purification. This process effectively compresses redundant intra-class variation while amplifying key discriminative features, endowing metacells with stronger transferability and prior representational capacity under few-shot learning scenarios. Extensive experiments on eight real-world single-cell transcriptomic datasets demonstrate that COMET significantly outperforms non-quantized few-shot annotation baselines under limited supervision, particularly in rare cell recognition tasks. Moreover, the purified metacell representations exhibit superior structural consistency in clustering evaluations compared with raw single-cell representations, validating the robustness and reliability of COMET in complex single-cell analysis settings. Code is available in https://github.com/lilyway18/COMET.

Physicochemically informed biological sequence generation has the potential to accelerate computer-aided cellular therapy, yet current models fail to jointly ensure novelty, diversity, and biophysical plausibility when designing variable regions of T-cell receptors (TCRs). We present PhysicoGPTCR, a large generative protein Transformer that is dual-conditioned on peptide and HLA context and trained to autoregressively synthesise TCR sequences while embedding residue-level physicochemical descriptors. The model is optimised on curated TCR-peptide-HLA triples with a maximum-likelihood objective and compared against ANN, GPTCR, LSTM, and VAE baselines. Across multiple neoantigen benchmarks, PhysicoGPTCR substantially improves edit-distance, similarity, and longest-common-subsequence scores, while populating a broader region of sequence space. Blind in-silico docking and structural modelling further reveal a higher proportion of binding-competent clones than the strongest baseline, validating the benefit of explicit context conditioning and physicochemical awareness. Experimental results demonstrate that dual-conditioned, physics-grounded generative modelling enables end-to-end design of functional TCR candidates, reducing the discovery timeline from months to minutes without sacrificing wet-lab verifiability.

Retrosynthesis identifies reactants to synthesize a target molecule, which is essential for drug discovery. Graph-edit-based methods transform products into reactants via molecular edits. While existing approaches reflect chemical transformations, they often model molecular structures and edit content separately. To bridge this gap, we propose RetroRecs, a graph-edit-based retrosynthesis model that unifies this encoding. RetroRecs builds a structured Reaction Context Sequence that unifies the dynamic encoding of intermediate molecules with the contextual embedding of each molecular edit. This design supports stepwise and correlation-aware retrosynthetic prediction. This linkage is established through two modules in each edit prediction step: the Reaction Context Reasoner (RCR), which analyzes relationships among reaction contexts (describing the "what, where, and when" of past edits), and the Molecular Environment Perceiver (MEP), which refines this contextual reasoning by incorporating the current synthon's structural features. Additionally, we propose a contrastive learning objective, guided by reaction types, to enhance the model's ability to differentiate between various edit process groups. These designs allow the model to make edit decisions that are sequentially coherent and thus chemically valid. Experiments on USPTO-50K show RetroRecs achieving 92.3% and 70.2% in top-10 exact match and round-trip accuracy, setting a new state-of-the-art for semi-template retrosynthesis models.

Quantifying macaque position, identity, and action is essential for neuroscience and pharmacology studies, yet remains challenging in cage environments due to frequent occlusions and rapid limb movements. This paper introduces the Primate Spatio-Temporal Relational Network (PSTRN), a framework designed to extract robust behavioral information from short video clips. PSTRN combines a hierarchical spatial encoder with a cross-frame relational attention module that applies temporal attention across all frames, adaptively aggregating complementary visual information to resolve ambiguities caused by occlusion or motion blur. A hierarchical 3D convolutional fusion then integrates multi-scale spatio-temporal features to jointly capture fine-grained limb motions and global body dynamics, enabling unified prediction of bounding boxes, identities, and actions. Experiments show that PSTRN achieves state-of-the-art performance, with 99.5% mAP50 and 85.1% F1 score on the single-macaque benchmark, and 92.6% mAP50 and 82.0% F1 score on the multi-macaque benchmark. These results demonstrate that PSTRN effectively captures macaque dynamics in realistic cage environments and offers a reliable solution for comprehensive behavior quantification.

Accurate segmentation of colorectal polyps and skin lesions is critical for clinical diagnosis. However, morphological variability and blurred lesion boundaries pose significant challenges to automated segmentation. Although the Mamba architecture excels in long-range modeling, its inherent 1D scanning mechanism often compromises 2D local geometric fidelity. Furthermore, conventional skip connections fail to adequately address semantic misalignment across different feature levels. To overcome these limitations, we propose DM-VMUNet, a hybrid architecture designed to integrate global context with local precision. Specifically, we introduce a Dual-Context Large-Small Convolution (DCLSConv) module. By leveraging multi-scale large-kernel perception and dynamic small-kernel aggregation, DCLSConv is designed to decouple receptive field expansion from computational cost, facilitating adaptive sharpening of lesion boundaries.Additionally, we propose a Multi-Scale Feature Alignment (MFA) module to semantically align encoder-decoder features prior to fusion. Extensive experiments on the Kvasir-SEG, ISIC 2017, and ISIC 2018 datasets demonstrate that DM-VMUNet achieves Dice coefficients of 90.75%, 89.87%, and 90.25%, respectively.These results validate the effectiveness of our approach, particularly its robust capability in delineating complex lesion boundaries compared to representative methods.

Given the scarcity of expert annotations, semi-supervised learning has become standard practice for medical image segmentation. The scarcity of expert annotations in medical imaging necessitates semi-supervised learning approaches. In semi-supervised medical image segmentation, existing methods typically impose point-wise consistency regularization or contrastive learning on individual voxels. However, such voxel-centric paradigm inherently lacks global structural constraints over class distributions, thus failing to disambiguate ambiguous boundaries with intra-class heterogeneity. While prototype learning provides distribution-level guidance, its semi-supervised application remains constrained by the reliability challenge of learning prototypes from noisy pseudo-labels. To this end, we present TR-PC, a framework that performs feature purification and semantic decoupling in probabilistic projection space. It first introduces a Texture-Refined Probabilistic Head (TR-Head) to suppress feature noise in 3D convolutions via local contextual modeling and dimension-balanced projection, thereby mapping raw voxels into high-fidelity multivariate Gaussian distributions. Afterwards, leveraging predictive uncertainty as a dynamic discriminator, it explicitly decouples each class center into cohesive core prototypes and dispersed boundary prototypes, and enforces divide-and-conquer likelihood matching. Therefore, the model is encouraged to infer and delineate highly challenging ambiguous regions via these decoupled core-boundary distributional conditions. TR-PC achieves state-of-the-art performance on three authoritative SSMIS benchmarks (LA, ACDC, and Pancreas-CT), with particularly significant improvements in boundary metrics (HD95).

Non-coding RNAs (ncRNAs) have been increasingly recognized as key regulators in diverse biological processes, and accumulating evidence suggests that interactions between ncRNAs and drugs play a critical role in disease regulation and therapeutic outcomes. Systematic identification of potential ncRNA–drug associations is therefore essential for understanding disease mechanisms and supporting drug discovery. Recent computational methods commonly integrate multiple similarity views of ncRNAs and drugs to improve prediction accuracy; however, heterogeneous biological similarity information often varies in reliability and exhibits inherent uncertainty, posing challenges for effective multi-view integration. To address this issue, we propose FCNDA, a novel ncRNA–drug association prediction model based on fuzzy consensus learning. FCNDA performs association prediction independently under each combination of ncRNA and drug similarity views and interprets view-specific prediction scores as fuzzy membership degrees, enabling explicit modeling of uncertainty and partial biological evidence. An adaptive fuzzy consensus coordination mechanism is then employed to aggregate multiple fuzzy memberships into a unified global prediction, where informative view combinations are emphasized while moderate discrepancies across views are tolerated. Experimental results on benchmark datasets demonstrate that FCNDA achieves competitive predictive performance and exhibits strong robustness to noisy and heterogeneous similarity information. Moreover, the learned consensus weights provide interpretable insights into the relative contributions of different biological similarity views.

Federated learning (FL) is severely impacted by model drift under data heterogeneity, manifesting as simultaneous feature representation drift and classifier drift across local clients—a dual challenge that collectively impairs global model aggregation performance. Current methodologies predominantly address a singular aspect of this challenge. In pursuit of a comprehensive resolution, we introduce a novel framework grounded in neural collapse theory. Our framework aims at achieving thorough drift mitigation by decoupling features from classifiers. We replace cross-entropy loss with hyperspherical uniformity gap loss in FL frameworks to mitigate classifier bias through data-agnostic means. Additionally, we reconstruct latent negative label knowledge distributions using knowledge distillation to ensure geometric consistency in feature representations across diverse clients. Extensive experiments validate the effectiveness of the proposed method in heterogeneous data.

As high quality public data becomes scarce, Federated Learning (FL) provides a vital pathway to leverage valuable private user data while preserving privacy. However, real-world client data often contains toxic or unsafe information. This leads to a critical issue we define as unintended data poisoning, which can severely damage the safety alignment of global models during federated alignment. To address this, we propose FedDetox, a robust framework tailored for Small Language Models (SLMs) on resource-constrained edge devices. We first employ knowledge distillation to transfer sophisticated safety alignment capabilities from large scale safety aligned teacher models into light weight student classifiers suitable for resource constrained edge devices. Specifically, during federated learning for human preference alignment, the edge client identifies unsafe samples at the source and replaces them with refusal templates, effectively transforming potential poisons into positive safety signals. Experiments demonstrate that our approach preserves model safety at a level comparable to centralized baselines without compromising general utility.

Federated learning, as an emerging distributed machine learning paradigm, offers a key advantage in effectively mitigating data silos while enhancing data privacy protection. However, heterogeneity in client data distributions often leads to insufficient generalization of the global model, thereby weakening personalized adaptation across clients. To address this issue, we propose a novel personalized federated learning method, FedPAL. On the server side, a class-conditional generator is constructed and trained using a classification loss and a supervised contrastive loss to generate feature distributions corresponding to class labels, so that the feature distributions exhibit intra-class compactness and inter-class separability. After downloading the generator, the client leverages the generated feature distributions to learn fusion weights between the local model head parameters and the global model head parameters. This process enables adaptive aggregation, which balances the absorption of global knowledge with the preservation of critical local information. As a result, an effective local model is initialized for each client at every communication round. To evaluate the effectiveness of FedPAL, extensive experiments are conducted on four benchmark datasets in the computer vision domain. The results demonstrate that the proposed method significantly outperforms ten state-of-the-art federated learning baseline methods.

Data heterogeneity (Non-IID) remains a critical challenge in federated learning (FL), often leading to model divergence and performance degradation. We propose TrustedMix, a novel framework that leverages Trusted Execution Environments (TEEs) to facilitate secure, privacy-preserving data sharing for distribution rectification. In TrustedMix, clients offload encrypted subsets of local data to a server-side TEE enclave, which computes the global label distribution and performs cross-client sample mixing within its hardware-isolated boundary. These synthesized samples are redistributed to clients to augment local training, effectively mitigating the impact of data skewness. Evaluated on CIFAR-10 and Tiny-ImageNet benchmarks, TrustedMix consistently outperforms state-of-the-art methods including FedAvg, FedProx, and FedMix. Specifically, in extreme non-IID scenarios ($\alpha=0.1$), the test accuracy of TrustedMix increases to up to 89.0\%, significantly surpassing the performance of FedMix while maintaining robust data confidentiality and minimal communication overhead.

Federated learning under client heterogeneity faces non-IID data distributions and client-specific feature statistics that degrade performance. While prior personalized federated learning methods partially mitigate statistical mismatch through local adaptation or normalization decoupling, they struggle to jointly model client-dependent discriminative cues and maintain cross-client semantic consistency. We propose FedPAF, a personalized federated framework that improves performance by jointly addressing normalization heterogeneity, feature heterogeneity, and semantic drift. FedPAF adopts a personalization strategy with normalization decoupling, leveraging Per-FedAvg-style meta-learning for adaptation while maintaining FedBN-style local batch normalization, enabling transferable representation learning while preserving client-specific statistics. It further highlights client-dependent discriminative cues through a lightweight multi-spectral channel reweighting module with negligible overhead, and aligns class semantics via global prototype aggregation to reduce semantic drift and enhance class-wise separability. Extensive experiments on Digits-Five, PACS, and DomainNet demonstrate consistent improvements over strong baselines in both accuracy and Macro-F1 under heterogeneous client settings.

Federated learning under non-independent and identically distributed (non-IID) data settings faces severe Byzantine security challenges. To mitigate the performance degradation of robust AGgregation Rules(AGRs) caused by gradient heterogeneity and the curse of dimensionality, a GrAdient Splitting (GAS) mechanism has been proposed. However, existing GAS method typically relies on a single reference when identifying Byzantine clients, making their defense performance sensitive to reference selection. This sensitivity arises because different attack types often require different reference choices to achieve satisfactory robustness. Motivated by this observation, we propose GAS-MR, namely Gradient Splitting with Multi-Reference Consistency-Based Byzantine Detection, which leverages multiple references to quantify the deviation-based inconsistency of Byzantine clients after gradient splitting, thereby enabling robust Byzantine client identification. Experimental results demonstrate that the proposed method consistently outperforms GAS under various nonIID data distributions and five representative Byzantine attack scenarios. Even in cases where performance is comparable, our approach exhibits superior robustness and stability, thereby significantly reducing the dependence on reference selection and prior knowledge of attack strategies Code is available at https://github.com/SKLIIS-AIS/GAS-MR.git.

Personalized federated learning (FL) tackles the challenge of data heterogeneity by training personalized models instead of a single global model. Backdoor attacks pose a fatal threat to FL systems. There is an urgent need to investigate on backdoor attacks under personalized FL. Existing representation-based personalized FL frameworks share a global extractor, which creates a large attack surface for backdoor attacks. This work explores previously unknown backdoor risks in FedPer and FedRep through data poisoning and parameter scaling. Experimental results show that the attack success rate (ASR) degrades dramatically under different pFL frameworks. We further propose a backdoor attack leveraging the shared representation extractor to inject backdoors more effectively. Experimental results demonstrate that our method improves the ASR by up to 72% while maintaining the ASR after stopping the attack for 200 rounds. Code is available at https://github.com/RezinChow/RBA.

Predicting the promoter methylation of O$^{6}$-methylguanine--DNA--methyltransferase (MGMT) from hematoxylin–eosin (H\&E)–stained glioma biopsies using deep learning (DL) remains challenging due to limited interpretability. In this work, we propose a weakly supervised framework that jointly performs classification and generates textual explanations grounded in visual evidence. We employ a lightweight convolutional neural network (CNN) to predict MGMT status from histology images. Building on these predictions, we introduce a Vision Transformer (ViT)–conditioned language decoder guided by SmoothGrad-CAM saliency maps to produce class-aware textual explanations. Patch-level predictions and saliency maps guide the generation of concise rationales aligned with visually relevant tissue regions. We further assess explanation faithfulness using perturbation-based insertion and deletion analyses. On a glioblastoma (GBM) cohort, the proposed model achieves a patient-level F1-score of 0.94, while the generated explanations remain consistent with both saliency maps and model predictions, demonstrating that class-conditioned textual rationales can provide a faithful and interpretable interface for histology-based MGMT prediction.

The rapid advancement of Visual Transformers (ViT) has created an urgent need for model interpretability, however, existing gradient-based explanation methods often suffer from significant visual noise caused by local gradient fluctuations. To address this challenge, this paper proposes the Gradient Local Mean-based (GLM) method, a novel framework designed to enhance the transparency of ViT models. The GLM method employs a stochastic sampling strategy by introducing Gaussian noise into the input image and aggregating the resulting Class Activation Maps (CAMs)—derived from gradient and attention matrices—to effectively mitigate noise instability. Extensive experiments on the ImageNet ILSVRC 2012 and CUB-200-2011 datasets demonstrate the method's superiority. On the fine-grained CUB-200-2011 dataset, the GLM method achieves a Top-1 Localization Intersection over Union (IoU) of 0.5958, significantly outperforming the state-of-the-art AGCAM method (0.4235) by 40.7%, while attaining a Pixel Accuracy of 0.8405. Furthermore, in pixel perturbation tests on ImageNet, the proposed method yields a superior Area Between Perturbation Curves (ABPC) score of 0.3739, statistically outperforming leading alternatives (p<0.05). These results confirm that the GLM method provides robust and precise visual explanations, facilitating reliable decision-making in practical applications.

Multimodal fake news detection faces two critical challenges: capturing fine-grained visual artifacts (e.g., splicing traces) and bridging the semantic gap between textual and visual modalities. Existing state-of-the-art approaches typically rely on standard Vision Transformers (ViT) or augment semantic information via automated image captioning. However, we argue that ViT's global attention mechanism often overlooks local discriminative details, while generated captions introduce semantic hallucinations and significant computational latency. To address these limitations, we propose SwinBERT-Fake, a novel framework that integrates Swin Transformer with BERT to prioritize hierarchical visual modeling over semantic augmentation. By leveraging Swin’s shifted window mechanism, our model effectively localizes subtle visual inconsistencies that standard ViT misses. Extensive experiments on the Fakeddit dataset demonstrate that SwinBERT-Fake achieves a state-of-the-art accuracy of 88.09%, significantly outperforming ViT-based baselines. Furthermore, our rigorous ablation studies reveal that semantic augmentation via captioning is redundant, offering no statistical performance gain while increasing inference latency by approximately 4.5 times. Finally, we introduce a comprehensive Explainable AI (XAI) module—incorporating Eigen-CAM, Integrated Gradients, and SHAP—to provide transparent visual, textual, and modal-level evidence for detection. Our findings suggest that optimizing the granularity of visual backbones is a more effective and efficient pathway for misinformation detection than relying on external generative models.

Reliable decision-making in medical applications has become increasingly crucial, particularly in skin lesion analysis, where misdiagnosis can lead to severe health consequences. While recent advances in computer vision have achieved high performance, relying solely on image-based models often overlooks essential clinical context. To address this limitation, we propose a novel multimodal framework that integrates dermoscopic image analysis with structured clinical metadata to improve both classification performance and decision reliability. The proposed approach combines an optimized Vision Transformer with a Deep Fuzzy Neural Network (ViT-DFNN) for image-based representation learning and a TabNet model for tabular data, enabling effective fusion of heterogeneous information through a decision-level weighted strategy. Experimental results on the HAM10000 dataset demonstrate the effectiveness of the proposed framework, achieving an accuracy of 93.15%, AUC of 96.1%, precision of 92.95%, sensitivity of 93.30%, specificity of 98.85%, and G-mean of 96%. These results confirm that integrating clinical metadata significantly enhances predictive performance and robustness compared to unimodal approaches. To ensure transparency, the framework incorporates Explainable Artificial Intelligence (XAI) techniques, where Score-CAM highlights important image regions and Global Feature Importance (GFI) identifies influential clinical variables. Overall, the proposed method provides a reliable and interpretable solution for AI-assisted skin lesion diagnosis.

Currently, datasets for Financial Sentiment Analysis (FSA) often contain rich multimodal information, yet the task-oriented evaluation metrics for processing this sentiment remain singular. Consequently, performance comparisons across different datasets are challenging, which hinders researchers from selecting appropriate resources for specific tasks. To address this problem, we propose FINS (Financial Integration Nexus of Standards), a data evaluation framework that systematically summarizes existing studies and abstracts the main application tasks of FSA into five core dimensions: cycle, risk, causality, entity and wording. On this basis, we establish a unified quantitative evaluation system for data. FINS combines large language models with financial toolchains to achieve an automated evaluation process, and presents the adaptability of data across different task dimensions through multi-dimensional quantitative score maps. Empirical results and real financial scenario experiment on several representative data sources show that FINS can effectively capture structural differences, which provides systematic support for researchers in data selection and modeling practices for specific tasks.

Alzheimer’s Disease (AD) staging from tabular clinical and neuropsychological data remains highly relevant for real-world decision support, yet model interpretability is hindered by multicollinearity among overlapping cognitive and functional measures. Using baseline Alzheimer’s Disease Neuroimaging Initiative (ADNI) data, we study one-vs-rest classification of Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD with linear and tree-based models: Elastic Net Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Beyond predictive performance, we provide a structured interpretability workflow that integrates Top-K global importance, cross-model feature agreement, correlation analysis of consensus predictors, and SHapley Additive exPlanations (SHAP). Results show a stable, clinically coherent core of predictors across models, with stage-specific shifts from global cognition and memory signals in CN to combined cognitive–functional impairment in MCI, and to severity and activities-of-daily-living measures in AD. Cross-model agreement highlights clinical dementia rating and functional measures as robust AD indicators. SHAP summaries further demonstrate consistent directional effects for key measures (e.g., higher CDRSB increasing the likelihood of AD and higher MMSE decreasing it). Overall, the proposed framework improves interpretability by contextualizing feature importance with cross-model robustness and correlation structure, supporting more reliable identification of AD-relevant predictors from tabular data.

Underwater images suffer from color distortion and detail loss due to wavelength-dependent absorption and scattering. Existing deep learning-based enhancement methods primarily operate in the pixel domain or rely on a single frequency-domain transform, limiting their ability to jointly preserve fine-grained details and model global dependencies. To address this gap, we propose WCI-Net, a wavelet-guided cross-domain interaction network for underwater image enhancement. WCI-Net decomposes features into multi-scale wavelet subbands to capture local structural details, while a cross-domain interaction module integrates wavelet-domain representations with Fourier-domain global refinement. A dual-domain loss further regularizes learning in both spatial and frequency domains. Extensive experiments on real-world benchmarks demonstrate that WCI-Net achieves state-of-the-art enhancement performance with low computational cost.

Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal: (1) reasoning-heavy models achieve peak utility in Free Mode via internal logic, whereas standard models require Strict Mode as a vital risk anchor; (2) alignment utility is regime-dependent, with Free and Guided modes capturing momentum in uptrending markets, while Strict Mode mitigates drawdowns in downtrends; (3) mid-scale models (35B) show optimal fidelity under strict constraints, whereas ultra-large models (122B) suffer an alignment tax under rigid rules but gain a performance premium in Guided Mode; (4) standard LLMs often fall into a high win-rate trap, optimizing for small gains at the expense of total returns, which can only be mitigated through deep reasoning or strict external guardrails. Project details are available at https://Strat-LLM.github.io.

Mixture-of-experts (MoE) language models promise favorable capacity–compute trade-offs, yet their routing decisions remain brittle: a single misrouted token can trigger compounding hallucinations or costly fallback heuristics, undermining reliability in real deployments. We propose RACER, a post-hoc, riskcontrolled routing layer that turns a black-box MoE (or any cascaded language system) into a selective generator: the system either (i) answers with a cheap expert path or (ii) escalates to a stronger expert (or retrieval-augmented path) when its own signals indicate elevated risk. RACERis built around conformal risk control and “learn-then-test” (LTT) calibration, yielding ffnitesample guarantees on selective risk under mild exchangeability assumptions. We evaluate RACERon two real tasks—TruthfulQA factuality and Natural Questions open-domain QA—using both a Mixtral-8x7B MoE system and a LLaMA-2-7B→70B cascade. Across settings, RACERsatisffes user-speciffed risk constraints while achieving 26–45% compute savings over always-large baselines at matched error rates, outperforming six competitive baselines including entropy thresholding, temperature scaling, and self-consistency routing. Ablations conffrm robustness to calibration set size (stable with n ≥ 150), distribution shift (topic/time splits), and score function choice. Code will be released upon acceptance.

Medical Visual Question Answering (MedVQA) supports clinical decision-making by answering diagnostic questions based on medical images. Multi-Question Learning (MQL) has been adopted to jointly model multiple questions per image, improving performance by leveraging inter-question dependencies. However, MQL also amplifies language bias, where models rely on textual patterns rather than visual content, by introducing a new shortcut: reasoning from correlated questions alone. Using only questions, MQL can exhibit severe language bias, yet achieves 84.35\% accuracy on the SLAKE dataset, nearly 30\% higher than traditional single-question learning methods, highlighting the risk of spurious shortcuts in reasoning. Further experiments conducted on the proposed SLAKE-CP-Co dataset, designed to isolate the language bias introduced by co-occurring questions, provide evidence for the presence of multi-question language bias in the MQL. Thus, this paper proposes a novel causal framework for MedVQA under MQL. The proposed method explicitly identifies two types of language bias: the conventional Single-Question Language Bias and the newly introduced Multi-Question Language Bias. Both are mitigated by subtracting their Natural Direct Effects from the total effect, enforcing greater reliance on visual information. Additionally, this paper introduces a Semantic-Aware Bias Subtraction module that dynamically adjusts the subtraction strength based on question semantics, preserving beneficial language cues for domain-knowledge questions. Experiments on SLAKE, VQA-RAD, SLAKE-CP and SLAKE-CP-Co demonstrate that the proposed method outperforms state-of-the-art biased and de-biased models in both accuracy and robustness. The proposed approach offers a principled, effective strategy to balance multi-question reasoning and language bias mitigation, advancing the state of the art in MedVQA.

To extend the benefits of dynamic lesion characterization to patients with contrast agent contraindications, synthesizing multi-phase contrast-enhanced computed tomography (CECT) from non-contrast CT (NCCT) has emerged as a vital tool for personalized radiotherapy planning. However, existing methods often struggle to fully utilize the anatomical context priors of NCCT and the temporal dependencies between different phases. This results in anatomical distortions across different phases and a significant increase in training iterations and parameter count, limiting their practical application. To overcome these challenges, this paper introduces a novel framework named RALSTM for the autoregressive and sequential synthesis of multi-phase CECT from abdominal NCCT. The framework leverages a Residual ConvLSTM module to model encoded features and capture the temporal transitions across phases. Additionally, a Region-Aware Gating Unit is used to localize contrast-enhancing regions and impose spatial constraints on the decoded features. Experimental results on both internal and external datasets demonstrate that this method achieves superior metrics and delivers outstanding synthesis quality.

Medical image segmentation remains challenging due to complex anatomy, ambiguous boundaries, and elongated structures, where structural inconsistency is frequently observed during feature reconstruction. In many encoder–decoder architectures, local upsampling operations are insufficient to preserve geometric continuity and long-range dependencies, resulting in fragmented predictions and boundary artifacts. To address this issue, a reconstruction-consistent segmentation framework, termed RKANet, is presented, in which decoding is explicitly strengthened as a structured reconstruction process. A cooperative modeling module is designed to jointly enforce local geometric refinement and global dependency propagation. In addition, a Boundary-Guided Skip Branch is incorporated to enhance skip feature fusion with boundary-aware shallow features. Extensive experiments on five public datasets demonstrate that RKANet achieves consistent performance improvements over representative methods, yielding a 3.00% Dice gain on BUSI and stable improvements of 0.25% – 0.55% on other benchmarks, with an average Dice score of 88.22%.

This paper proposes an explainable machine learning framework to support the diagnosis of urinary tract diseases, particularly bladder cancer, by employing SHAP (SHapley Additive exPlanations)-based feature selection to enhance model transparency and predictive performance. Six binary classification scenarios were designed to differentiate bladder cancer from other urological and oncological conditions. The models were implemented using XGBoost, LightGBM, and CatBoost algorithms, with hyperparameter optimization conducted via Optuna and class imbalance addressed through the SMOTE technique. Predictive variables were selected according to their importance values derived from SHAP analyses. The results demonstrated that SHAP-based feature selection efficiently reduced data dimensionality while maintaining or even improving performance metrics such as balanced accuracy, precision, and specificity. These findings underscore the potential of explainable artificial intelligence in medical diagnostics. The proposed methodology contributes to the development of transparent, reliable, and efficient clinical decision support systems, helping to optimize the screening and early diagnosis of urinary tract diseases, particularly bladder cancer.

Asynchronous Federated Learning (AFL) enables distributed model training without client synchronization, offering enhanced scalability in heterogeneous and bandwidth-limited environments. However, in many AFL approaches, the use of gradient compression is coupled with fixed aggregation intervals and compression rates, which often leads to suboptimal performance caused by excessive model staleness and communication cost. We propose FedJACA, a novel AFL framework that jointly and adaptively configures the aggregation interval and gradient compression rate, so as to maximize gradient utility while minimizing staleness and communication cost. We model their intertwined impact and formulate a joint optimization problem under training time and bandwidth constraints. And we design a customized solution algorithm based on the Trust Region Newton method that yields high-quality solutions for achieving a balanced trade-off between efficiency and cost. Extensive experiments on four benchmarks show that FedJACA improves accuracy, reduces training time by 40%, and lowers communication cost by 35%, outperforming state-of-the-art baselines.

Diffusion models have demonstrated powerful performance in generating high-quality images. A typical example is text-to-image generator like Stable Diffusion. However, their widespread use also poses potential privacy risks. A key concern is membership inference attacks, which attempt to determine whether a particular data sample was used in the model training process. Existing membership inference attacks against diffusion models either directly exploit sample loss differences or rely on image-level reconstruction differences. Both approaches commonly ignore the consistency characteristics of noise prediction during the diffusion process, resulting in either low inference accuracy or high computational costs. To address these shortcomings, we propose a membership inference method based on noise aggregation analysis, and introduce a single-step, low-intensity noise injection diffusion strategy to amplify differences between member and non-member samples. Our proposed approach substantially reduces model query requirements while delivering more efficient and accurate membership inference.

To ensure secure face privacy protection, existing class-specific adversarial methods have shown substantial performance improvement by tailoring perturbations to individual identities. However, due to high local similarity in same-identity images, these methods focus perturbations on specific regions. Though some studies use generic features, they still fail to adequately balance generic and fine-grained features, yielding limited privacy protection improvements. To address these issues, we innovatively propose Boosting Class-specific Face Privacy Protection with Stochastic Mask Regularized Adaptive Momentum (SMRAM) for privacy protection. This approach introduces Stochastic Mask Regularized Adaptive Momentum, which dynamically integrates fine-grained features and global consistency features through Adaptive Momentum (AM). Additionally, Stochastic Mask Regularization (SMR) increases the diversity of the perturbation distribution, thereby capturing richer generic features. When fine-grained features dominate, AM refines the generation direction by incorporating generic features and stabilizes the generation process through momentum accumulation, thereby improving the success rate of privacy protection. Furthermore, SMR forces the model to learn more generic features, enhancing the generalization of the perturbations. Extensive experiments conducted on two benchmark datasets demonstrate that the proposed SMRAM outperforms the state-of-the-art approaches, achieving a higher success rate in privacy protection for same-identity images under black-box recognition models.

Semi-supervised semantic segmentation has made significant advances, yet current methods often overlook variations in sample difficulty and rely on limited data augmentation strategies, which constrains both accuracy and the exploration of diverse perturbation spaces. This paper proposes a Dual Adaptive Augmentation framework. First, we introduce a quantitative hardness evaluation mechanism based on the Intersection over Union (IoU) of segmentation predictions, which assesses the learning difficulty of each unlabeled sample. Guided by the derived hardness coefficient, we dynamically adjust the mixing ratio of weak and strong augmentations, aligning perturbation intensity with the model's evolving learning state. Second, we expand the perturbation space by applying two independent strong augmentation strategies to unlabeled images, thereby improving generalization. Finally, we introduce additional random perturbations, such as Dropout2D, between the encoder and decoder layers to enhance feature-level diversity and robustness. Extensive experiments demonstrate that our model consistently outperforms existing methods across different data partition settings.

Multi-physics simulations play an essential role in accurately modeling complex interactions across diverse scientific and engineering domains. Although neural operators, particularly the Fourier Neural Operator (FNO), have significantly improved computational efficiency, they often fail to capture the intricate correlations inherent in coupled physical processes. To address this limitation, we introduce COMPOL, a novel coupled multi-physics operator learning framework. COMPOL extends conventional operator architectures by incorporating a sophisticated attention-based aggregation mechanism that effectively models interdependencies among interacting physical processes within latent feature spaces. Our approach is architectureagnostic and seamlessly integrates into various neural operator frameworks involving latent space transformations. Extensive experiments on diverse benchmarks, including biological reactiondiffusion systems, pattern-forming chemical reactions, multiphase geological flows, and thermo-hydro-mechanical processes, demonstrate that COMPOL consistently achieves superior predictive accuracy compared to state-of-the-art methods. Our code and data are available at https://github.com/AriaQJ/COMPOL.

This paper introduces GAF-Flow, a novel auditory-perceptual flow model for high-fidelity neural Text-To-Speech (TTS) synthesis. To address the limitations of “machine-centric” acoustic modeling, which often overlooks the non-linear frequency selectivity of human hearing, we propose three key innovations. First, a Gammatone-Mel Hybrid Filterbank is designed to mimic cochlear resolution and enrich high-frequency spectral details. Second, a Channel-wise Adaptive Normalization module harmonizes the heterogeneous statistical distributions between filter domains. Third, an auditory-prior supervision scheme provides targeted perceptual guidance during training. Experimental results on LJSpeech and VCTK datasets demonstrate that GAF-Flow significantly outperforms state-of-the-art baselines in both subjective (MOS, SMOS) and objective metrics (MCD, PESQ), successfully restoring fine spectral details and reducing artifacts while maintaining efficient inference. Our work bridges the gap between physiological auditory mechanisms and generative speech modeling, providing a perceptually-enhanced paradigm for high-quality synthesis.

Sea Surface Temperature (SST) is a key factor in ocean-atmosphere interactions, which exerts significant influence on climate, ecology and maritime activities. Existing methods are limited to the modeling and optimization of long-term temporal dependence, and fail to adequately capture the complex temporal patterns in SST sequences, which are formed by the combined effects of long-term periodic trends and short-term non-stationary variations, resulting in suboptimal performance for long-term SST prediction tasks. To address this issue, we propose a temporal periodic feature mining and reconstruction transformer (TPRformer). In TPRformer, the encoder employs a temporal decomposition block to separate the historical sequence into periodic and non-stationary components, with a focus on mining long-term temporal dependencies within the periodic component to construct periodic features. The decoder performs the temporal decomposition on the SST sequence, employs temporal cross-attention to dynamically align historical periodic features, and ultimately achieves SST prediction by reconstructing the two decomposed components. Experimental results on multiple SST datasets demonstrate that TPRformer is superior to state-of-the-art baseline methods across multiple forecasting horizons.

Unsigned Distance Functions (UDFs) provide flexible representations for non-watertight geometries but struggle with a critical spectral mismatch. Since 3D surfaces exhibit non-stationary frequency distributions, demanding high frequencies for sharp edges but low frequencies for flat regions, traditional global encodings are inherently inefficient. They typically enforce excessive bandwidth globally to resolve local details, inevitably inducing computational redundancy and ripple artifacts in smooth areas. To address this, we propose a geometry-aware Learnable Gabor Network. First, leveraging the time-frequency locality of Gabor wavelets, our network adaptively modulates its spectral bandwidth, concentrating representational capacity on geometric details while preventing noise in flat regions. Second, to mitigate the gradient ambiguity of UDFs, we introduce a Dual-Query Geometric Consistency mechanism. By imposing local differential constraints, we regularize the gradient field direction without relying on ground-truth normal supervision. Experiments demonstrate that our approach improves upon state-of-the-art methods, achieving a superior balance between reconstruction fidelity and efficiency.

Interactive Imitation Learning (IIL) is a promising paradigm for the online optimization of robotic skills, enabling agents to learn from sparse expert demonstrations. However, in few-shot scenarios, effectively combining human feedback with data aggregation to mitigate distributional shift remains a significant challenge. While previous works explored this integration, systematic research into internal mechanisms—such as dynamic feedback weighting, delay compensation, and lightweight trajectory resampling—remains insufficient under extreme data constraints. This paper proposes the Agile Iterative Aggregation (AIA) framework, which systematically optimizes these core iterative feedback and data processing mechanisms using a human-in-the-loop approach. We evaluate AIA on six simulated manipulation tasks and further validate it on a real-world platform. The results show that our optimized processing and aggregation strategies significantly enhance policy performance and robustness. Compared to baselines, AIA achieves superior success rates and finer sub-skill acquisition, offering a cost-effective solution for high-efficiency robotic skill learning.

Generating long, realistic, and semantically aligned full-body human--object interactions from natural language remains challenging due to complex hand--object contacts, long-range temporal dependencies, and multi-part coordination. Existing methods often struggle to jointly model body motion, hand grasping, and object dynamics. To address these challenges, we propose a language-guided framework for full-body grasp motion generation with explicit contact awareness and part-wise coordination. We introduce a unified representation that encodes SMPL-X body motion, object motion, contact labels, and relative spatial features, and decompose interaction sequences into four semantic components that are discretized using component-specific VQ-VAEs. A masked autoregressive Transformer with a part coordination mechanism is then employed to generate coherent motion tokens conditioned on text. Experiments on the GRAB dataset demonstrate that our method generates longer and more coordinated interaction sequences, outperforming existing approaches. The code will be made publicly available upon acceptance.

Long-term multivariate time series forecasting (LTSF) demands accurate long-horizon prediction while modeling cross-variable dependencies. Most Transformer-based methods focus attention along the temporal axis, underutilizing inter-variable interactions and suffering from error accumulation over long horizons. We propose DualMixFormer, a variable-centric framework that treats each variable as a token and applies self-attention across variables. To enhance token representations amid heterogeneous periodicity, we introduce MEVA, a multi-embedding module that injects channel identity and phase embeddings to guide cross-variable attention explicitly. For robustness, DualMixFormer employs a dual-branch head: a lightweight MLP branch provides a low-variance continuation bias for smooth components, while a MEVA-conditioned Transformer branch models nonlinear residuals. Predictions are fused via learnable channel-wise weights, adaptively combining smooth and residual components. RevIN is integrated to address instance-wise distribution shifts. Experiments on seven benchmarks show competitive performance across horizons. Ablations and visualizations confirm the model’s effectiveness, interpretability, and robustness.

Non-contact cattle face recognition is the prerequisite for precise livestock management. However, modern intensive farming systems predominantly employ frozen semen technology for reproduction, which leads to highly similar genetic characteristics among individuals (particularly half-siblings sharing the same paternal lineage), presenting a significant challenge for individual identification. To address this issue, this paper proposes the DiffSiam network, which incorporates a Difference-Aware Residual Attention Module (DA-RAM). This approach models the recognition task as a fine-grained similarity measurement for image pairs. Specifically, the Swin Transformer V2 is employed as the core architecture for deep feature extraction. Subsequently, the DA-RAM module is utilized to explicitly compute feature difference maps, which guide the network to focus on highly distinctive local discrepancies, such as facial texture and contours, facilitating identity determination based on pairwise similarity scores. Additionally, the P-K sampling strategy combined with a weighted binary cross-entropy loss is employed to optimize the distribution of positive and negative samples. Experimental results on a dataset comprising 198 categories and 27,118 genetically similar cattle faces demonstrate that the proposed method achieved 98.09% Top-1 accuracy and 98.12% F1 score, outperforming existing mainstream approaches. The ablation experiments further confirmed that the DA-RAM module delivered a 15.04% performance improvement, which demonstrates its efficacy for fine-grained recognition.

Large Neighborhood Search (LNS) and 2-opt local search are key algorithms for routing problems like TSP and CVRP. Standard hybrid frameworks usually separate these steps, using LNS for global exploration and 2-opt for later optimization. This order often leads to fault propagation, where suboptimal edges from repair are only fixed after costly global scans. To address this limitation, we investigate a \textit{Process-Aware Window Mechanism} for 2-opt based LNS. Unlike other granular search methods that use fixed geometric metrics, our method uses the dynamic information from the repair operator. We set an optimization 'window' around the insertion index of new nodes. This change makes 2-opt a proactive, built-in correction tool instead of just a post-processing step. We test six window selection strategies, from static to adaptive. Our experimental analysis shows that this frequent, low-complexity micro-optimization can improve convergence speed and result stability compared to traditional global search methods.

Abstract— This work presents a mathematically grounded, non-reductionist framework for simulating emergent dynamics in a canonical 10-dimensional ancient medicine-artificial immune inspired system . The system is realized as a stochastic dynamical network with state-dependent drift, adaptive control operators, and hysteresis-based memory, enabling trajectory-level exploration of homeodynamic plasticity. Interventions are modeled as smooth, time-gated operators deforming the state-space geometry without imposing fixed-point constraints. Simulations demonstrate separable pathological regimes, regime-specific stability, and adaptive recovery under structured interventions, illustrating the system’s intrinsic self-regulation, non-commutativity, and memory effects. This framework unifies emergent intelligence analysis, trajectory-based disease modeling, and intervention dynamics in a fully differentiable, reproducible computational pipeline, providing a platform for exploring complex adaptive systems beyond traditional reductionist approaches

Earth observation satellites have been widely applied and become indispensable in various domains. The collaboration of optical, synthetic aperture radar, and electromagnetic satellites enables diverse mission requirements through complementary capabilities. However, scheduling observation tasks across multiple satellite types to satisfy diverse objectives still remains a critical challenge, especially on large-scale instances. This paper focuses on two optimization objectives, i.e., maximizing the total profit of scheduled tasks and maximizing the average profit efficiency across satellites. To solve this problem, a dual-layer clustering guided genetic algorithm is proposed to obtain high-quality and well-distributed solution sets. The algorithm decouples the scheduling process into two layers, i.e., task sequence construction and time window assignment, and adopts dual sequence encoding. The two layers dynamically switch based on the performance improvement to generate new task sequences or new time-window assignments. At both of task and time-window layers, clustering is performed to guide the crossover and mutation for improving profits and resource utilization based on task characteristics and historical selection frequencies. In addition, a time-window allocation strategy is adopted in both layers to improve solution quality by utilizing historical information and current resource states. Experimental results on fifteen instances with up to 1300 tasks demonstrate that the proposed algorithm outperforms state-of-the-art algorithms in terms of convergence and diversity.

The design of conformal cooling channels (CCC) in injection molding is a complex, multi-objective engineering task in which geometric, thermal, and mechanical constraints interact in nonlinear ways. Although evolutionary multi-objective optimization (EMO) algorithms, such as NSGA-II, have been successfully applied to this domain, the interpretation of large Pareto-optimal sets often relies on manual visual inspection, which limits knowledge transfer and decision reproducibility. This paper examines how interactive knowledge discovery, facilitated by the Mimer platform, can enable the systematic interpretation of EMO results in the injection molding process. Four bi-objective CCC optimization problems were solved using high-fidelity simulations accelerated by an Artificial Neural Network surrogate model based on Moldex3D data. Mimer was then used to analyze the resulting solution sets through linked visualizations, region-of-interest-based preference articulation, and flexible pattern mining for rule extraction. The generated knowledge was compared against the independent assessment of an expert in mold design. Results show that Mimer consistently complements expert insights, identifies the most influential geometric variables, and uncovers additional non-obvious relations across runs, including stable structural rules governing channel diameter, distance to the cavity surface, and gate-dependent performance. These findings demonstrate that Mimer can act as a semi-automatic analyst in EMO-based CCC design, increasing transparency, supporting informed decision-making, and improving the reusability of optimization knowledge in industrial injection molding.

Graph neural networks (GNNs) have achieved great success in many applications. Most existing GNNs rely on backpropagation for end-to-end training. Recently, forward-only learning methods have been considered as promising alternatives, but directly applying them to GNNs faces challenges in efficiency and training stability. In this paper, we propose the Prototype-based Forward algorithm for Graph Neural Networks (ProF-GNN), a forward-style GNN learning framework that introduces class-aware virtual nodes to explicitly model class prototypes. By aligning node representations with their corresponding prototypes at each layer, ProF-GNN enables direct and efficient prediction without exhaustive label-wise inference or the generation of multiple negative samples. Furthermore, we incorporate a diversity-aware objective to prevent prototype collapse, ensuring stable and discriminative layer-wise learning. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed learning framework for GNNs for node classification.

Graph Neural Networks (GNNs) have shown remarkable effectiveness in learning graph representations across diverse applications. However, in many real-world scenarios, graph-structured data often contains noisy labels, which substantially compromise the performance of GNNs. This degradation occurs because noisy labels weaken effective supervision during training, causing GNN-based models to memorize mislabeled samples and consequently impair their performance on downstream tasks. Existing approaches for handling label noise in images and text typically rely on large-scale labeled datasets and assume data independence, making them unsuitable for addressing the label sparsity and structural dependencies inherent in graphs. To overcome these challenges, we propose a multi-network learning paradigm in which models supervise each other, thereby reducing the bias accumulated when learning from noisy samples. We introduce a novel multi-network framework, Joint Training with Consistency and Diversity (JoCaD), which is designed to maximize prediction consistency across networks while maintaining sufficient diversity in their representation learning. Specifically, we incorporate graph augmentation through link addition to enrich the aggregated information, and employ a consistency loss to ensure alignment between the node representations learned by both networks. Extensive experiments on four real-world datasets demonstrate the effectiveness of the proposed JoCaD in addressing the node classification task under varying levels of label noise.

Heterogeneous graphs provide a representation for real-world systems that involve multiple entity types and relations. In many of these applications, the practical goal is to detect one class of interest while treating the remaining instances as outliers, which naturally leads to one-class learning on graph neural networks (GNNs). However, most one-class GNN approaches either rely on unsupervised methods, decouple representation learning from the one-class objective in two-step pipelines, or focus on homogeneous graphs and do not explicitly leverage heterogeneity in generic settings. We propose FOLGA (Fusion-rewired One-cLass Graph Autoencoder), an end-to-end method for one-class node classification on heterogeneous graphs. FOLGA enriches heterogeneous structure through complementary rewiring strategies, fuses information from multiple rewired graph views via a fusion GNN encoder, and combines graph reconstruction with a state-of-the-art sphere loss to learn interpretable 3D embeddings. We also show that FOLGA naturally extends to dynamic data streams for evolving dynamic heterogeneous graphs in online learning. Experiments with 12 representative baselines in three datasets demonstrate that FOLGA outperforms them, while providing real-time interpretability of the learning dynamics and competitive results.