Neural networks have been widely used in facial expression recognition (FER) because they extract image features automatically and have achieved significant results. Nevertheless, manually designing neural networks is prone to errors and consumes a significant amount of time, while automatically designing neural networks suffers from high time consumption and large memory occupation. To address the above problems, we propose a gradient-based lightweight network automated design method called GLA-FERNet for FER. First, we propose a gradient FER network framework based on differentiable neural architecture search to achieve accurate FER and reduce time consumption during automated search of network architectures. Second, we propose a channel sampling process strategy by random or sequential sampling a small part of the super-network to enhance search stability while maintaining low memory usage. We conduct rich experiments on RAF-DB, AffectNet, RAF-DB (10 %), RAF-DB (20%) and RAF-DB (30%) datasets to verify the performance of the proposed GLA-FERNet. The experimental results demonstrate that GLA-FERNet outperforms state-of-the-art representative methods in terms of accuracy, precision, recall, and F1-score on the five benchmark datasets. Furthermore, the results of ablation experiments demonstrate that the proposed channel sampling process strategy greatly improves the performance of the designed neural network and significantly reduces search time for large datasets.

Continual Learning (CL) aims to develop AI models that learn effectively from sequential tasks while mitigating catastrophic forgetting. Replay-based methods have emerged as a promising solution for CL, which stores a subset of past exemplars and then replays it to preserve prior knowledge. Existing exemplar selection strategies predominantly focus on feature-space representativeness but overlook output distribution variation. In this work, we identify that neighboring samples in feature space may sustain significantly different output probability distributions. This indicates that the nearest neighbors to class-wise mean feature vectors do not consistently serve as optimal representative samples. We further demonstrate that predictive uncertainty serves as a reliable indicator of such non-representative samples. Building on this insight, we propose Predictive Uncertainty Controller (PUC), which aims to benefit replay-based CL methods by filtering out samples with excessive uncertainty. Extensive experiments validate our approach, showing that PUC consistently enhances CL performance when integrated with existing replay-based methods.

Neural architecture search (NAS) has been widely adopted to design high-accuracy architectures, which are often vulnerable against adversarial attacks. To address this problem, existing robust NAS methods mainly focus on optimizing both natural accuracy and adversarial robustness in a fixed supernet, which is designed for natural accuracy. As a result, the derived architectures have the same construction scheme as the supernet, suffering from limited adversarial robustness and flexibility. In this article, we present the ARNAS++ method to search for accurate and robust neural architectures via a supernet with flexible parameter budgets and width. Specifically, we propose a parameter budget controlling loss to make architectures contain less parameters in the rear cells, based on which the adversarial robustness can be guaranteed. Moreover, we also propose a learnable filter number reduction ratio to control the filter numbers in the supernet, which can find more robust architectures beyond the fixed supernet, and make the supernet more flexible at the same time. We conduct experiments on six widely used benchmark datasets against the state of the art. The experimental results demonstrate that the proposed ARNAS++ method outperforms the competitors in terms of both natural accuracy and adversarial robustness under various popular adversarial attacks. In addition, the ablation studies show the effectiveness of the designed components and their positive contributions to the overall performance. The source code is available at: https://github.com/fyqsama/ARNASpp

Spiking Large Language Models (LLMs) have emerged as an energy-efficient alternative to conventional LLMs through their event-driven computation. To effectively obtain spiking LLMs, researchers develop different ANN-to-SNN conversion methods by leveraging pre-trained ANN parameters while inheriting the energy efficiency of SNN. However, existing conversion methods struggle with extreme activation outliers and incompatible nonlinear operations of ANN-based LLMs. To address this, we propose a loss-less ANN-SNN conversion for fully spike-driven LLMs, termed LAS. Specifically, LAS introduces two novel neurons to convert the activation outlier and nonlinear operation of ANN-based LLMs. Moreover, LAS tailors the spike-equivalent Transformer components for spiking LLMs, which can ensure full spiking conversion without any loss of performance. Experimental results on six language models and two vision-language models demonstrate that LAS achieves loss-less conversion. Notably, on OPT-66B, LAS even improves the accuracy of 2% on the WSC task. In addition, the parameter and ablation studies further verify the effectiveness of LAS.

OBJECTIVES:This study aims to compare the performance of different large language models (LLMs) in pre-exam learning for the Chinese dental licensing examination, with a focus on evaluating their differences in answering, explanation, and teaching effectiveness, to provide a reference for the application of LLMs in dental education. METHODS:Three evaluation scenarios were designed: selecting correct answers, providing answer explanations, and adversarial testing. DeepSeek-R1, Qwen 2.5-MAX, Doubao 1.5 Pro, Xinghuo Spark-X1, ERNIE 4.0 Turbo, GPT-4o, and Huaxi Zhilian were selected for comparative testing. Evaluation metrics included accuracy, net accuracy, and pedagogical effectiveness. RESULTS:In the scenario of selecting correct answers, all LLMs exceeded the passing threshold, with Huaxi Zhilian achieving the highest accuracy (84%). In the answer explanation scenario, Huaxi Zhilian demonstra-ted the highest net accuracy (92%), followed by DeepSeek-R1 (89%), among models. Regarding pedagogical effectiveness, Huaxi Zhilian ranked highest in relevance, practicality, and clarity, whereas GPT-4o led in conciseness, among the investigated LLMs. In adversarial testing, Huaxi Zhilian and DeepSeek-R1 exhibited the smallest declines in accuracy and net accuracy, respectively, among the tested models. CONCLUSIONS:In pre-exam learning for the Chinese dental licensing examination, knowledge-enhanced LLMs specifically optimized for dentistry (e.g., Huaxi Zhilian) outperform reasoning LLMs pretrained on general Chinese corpora (e.g., DeepSeek-R1) and those primarily trained on English corpora (e.g., GPT-4o). However, the performance of all models declines under adversarial conditions. Future research should focus on addressing identified weaknesses to enhance the utility of LLMs in dental education further.

Zero-shot neural architecture search offers a training-free paradigm for automated architecture design by employing zero-cost proxies. Recently, the minimum eigenvalue of correlation (MeCo) on feature maps has been identified as a strong gradient-free indicator of architecture performance. However, MeCo computation relies on the feature maps of each and every layer in the network, resulting in significant computational overhead. In this work, we reveal the computational redundancy of MeCo, showing that an accurate assessment of network capacity can be achieved without feature maps of all layers. Strengthened by a theoretical analysis, we propose Faster MeCo (FMeCo), which focuses exclusively on activation features for MeCo calculation. In addition, we integrate a simplified scoring method for architecture’s input discrimination capability, also based on activation features, into FMeCo, to create FMeCo-SWOT. Experimental results on various NAS benchmarks, datasets, and search spaces demonstrate the superiority and efficiency of our proposed proxies. Notably, both FMeCo and FMeCo-SWOT take merely 1/3 of the time required by MeCo for evaluating 1000 architectures in NAS-Bench-201 with higher accuracy. Relying only on activation features, our proposed proxies also achieve strong performance in evaluating both CNN and ViT architectures.

Accurately reconstructing three-dimensional urban wind and pollutant concentration fields from sparse observations is critical for emergency response, air quality management, and urban planning, yet remains computationally prohibitive with direct CFD approaches. This study introduces the Recurrent Encoder–Proper Orthogonal Decomposition Decoder (RE-POD) framework, combining POD dimensionality reduction with recurrent neural network encoders to map sparse sensor time series directly to the POD latent space, thus enabling simultaneous field reconstruction without requiring known boundary conditions at inference time. Four encoder architectures are compared on a 200-scenario large eddy simulation ensemble of the MUST experiment with 1.88 million mesh nodes and 93 virtual probes: SHRED-POD (LSTM baseline), Transformer-POD (multi-head self-attention), DualDecoder-POD (task-separated decoding heads), and CNN-GRU-POD (hierarchical spatial-temporal encoder). CNN-GRU-POD achieves the strongest sparse-sensor performance with FAC2 = 0.697, FMS(1.0 ppm) = 0.764, VG = 2.869, and NMSE = 6.329. DualDecoder-POD emerges as a strong candidate for wind field reconstruction, delivering the lowest concentration VG (2.480) and NMSE (5.870) under dense-sensor configurations. Vertical profile analysis reveals concentration accuracy degrades near the surface (Z < 2.5 m), whilst horizontal wind components maintain R² > 0.96 above this height and the vertical component achieves R² = 0.72–0.92, with all components degrading with height. Sensitivity analysis over 80 configurations shows reconstruction quality stabilizes beyond 100 training snapshots and 10–20 probes, and robustness tests across 50 random probe placement scenarios confirm that reconstruction quality is largely insensitive to sensor positioning. These findings provide guidelines for deploying RE-POD-based surrogate systems in practical urban environmental monitoring.

Balancing convergence, diversity, and feasibility is a critical challenge in constrained multi-objective optimization, particularly when dealing with discontinuous or degenerate Pareto fronts (PFs). Constrained multi-objective evolutionary algorithms (CMOEAs) are commonly employed to tackle such problems, yet most of them prioritize convergence and feasibility while overlooking diversity. To address this gap, we propose a novel two-stage hybrid tri-population CMOEA (TS-HC3) that enhances both exploration and exploitation to achieve well-converged, diverse, and feasible solutions. In the first stage, a dual-population exploration mechanism is presented to promotes broad exploration and accelerate the discovery of promising feasible regions. In the second stage, a tri-population exploitation mechanism is introduced, incorporating a novel subregion-aware environmental selection strategy that adaptively preserves solution distribution and uniformity across the PF. A two-level multi-armed bandit strategy is proposed to dynamically select the most suitable parent population and variation operator based on real-time evolutionary feedback. Finally, a differential evolution based flexible genetic operator is designed to enhance exploration in sparse feasible regions. Extensive experiments on widely used benchmarks demonstrate that TS-HC3 significantly outperforms state-of-the-art CMOEAs, particularly on problems with discontinuous or degenerate PFs. The effectiveness of TS-HC3 is further validated on a multi-unmanned surface vehicle collaborative path planning problem, highlighting its practical applicability in complex real-world scenarios. Moreover, an ablation study is conducted to validate the contribution of each individual component within TS-HC3.

Spiking neural networks (SNNs) have the merit of energy efficiency and have been widely used for various real-world applications. Similar to other types of neural networks, the performance of SNN is also significantly decided by its architecture. In this article, we propose an evolutionary multiobjective spiking neural architecture search (EMO-SNAS) method that completely enables the automatic design of SNN architectures with both high performance and low power consumption. To achieve this, we first design a variable-length encoding strategy for SNNs, addressing the issue that traditional encoding strategies need to manually set the depth in advance. Furthermore, we propose an exploitation operator focusing on the local search for the variable-length encoding, as well as an exploration operator focusing on the global search based on the temporal expansion. Based on NSGA-II, EMO-SNAS can greatly balance the performance and power consumption during the architecture design. Experiments on three widely used image classification datasets show that EMO-SNAS can achieve the best among the state-of-the-art methods. Specifically, EMO-SNAS gains 0.45%, 0.26%, and 7.52% in terms of classification accuracy, yet significantly contributes to 33%, 29%, and 12% fewer spike numbers on CIFAR10, CIFAR100, and TinyImageNet datasets. Ablation studies show that the temporal expansion can improve the performance of EMO-SNAS. Moreover, the measurement of power consumption and theoretical convergence of EMO-SNAS are also discussed to justify its component design. In addition, with EMO-SNAS, the impact of initial channels for SNNs is also systemically investigated, based on which a conclusion against existing consensus is achieved. The source code is available at https://github.com/songxt3/EMO-SNAS.

Neural architecture search (NAS) aims to automatically design high-performance architectures of deep neural networks, which have shown great potential in various fields. However, the search process of NAS is computationally expensive since plenty of deep neural networks are trained to get the performance on GPUs. Performance predictors can directly estimate the performance of architectures without GPU-based training, thus can overcome this barrier. However, the construction of performance predictors requires labeling plenty of architectures sampled from the corresponding NAS search space, which is still prohibitively costly. In this paper, we propose a Domain Adaptive performance Predictor (DAP), which can construct a performance predictor based on the labeled architectures provided by existing benchmarks and then enable it to other search spaces via domain adaptive techniques. To achieve this, we first propose a domain-agnostic feature extraction method to refine the domain-invariant features of neural architectures. Then, we propose a novel embedding method to learn the shared representations of architecture operations. Experimental results demonstrate that DAP outperforms eight baselines upon six popular search spaces. Notably, we only require the search cost of 0.0002 GPU Days to find the architecture with 77.10% top-1 accuracy on ImageNet and 97.86% on CIFAR-10. In addition, we show the theoretical upper bound of the generalization error in the target search space, further illustrating the generalizability of DAP. The source code is available at https://github.com/songxt3/DAP.

Evolutionary neural architecture search (ENAS) is an important part of evolutionary machine learning, and has empirically achieved great success. However, its rigorous theoretical analysis has yet to be touched. This work presents preliminary steps toward the mathematical runtime analysis of ENAS. In particular, we define a binary classification problem UNIFORM, and formulate an explicit fitness function to represent the relationship between neural architecture and classification accuracy. Furthermore, we consider a (1+1)-ENAS algorithm and (μ+λ)-ENAS algorithm to find an optimal neural architecture capable of solving Uniform. We theoretically present their expected runtime and prove that: 1) the (1+1)-ENAS algorithm with local mutation optimizes the Uniform problem in linear time and global mutation does not make the algorithm faster; 2) the (μ+λ)-ENAS algorithm is not asymptotically faster than the (1+1)-ENAS algorithm when the offspring population size λ ∈ ω(ln n/(ln ln n)), where n is the UNIFORM problem size; 3) for constant λ, the (μ+λ)-ENAS algorithm with global mutation outperforms the local mutation version in terms of the lower bound of expected runtime. Empirical studies also confirm the theoretical results.

The adaptation of large-scale Vision-Language Models (VLMs) through post-training reveals a pronounced generalization gap: models fine-tuned with Reinforcement Learning (RL) consistently achieve superior out-of-distribution (OOD) performance compared to those trained with Supervised Fine-Tuning (SFT). This paper posits a data-centric explanation for this phenomenon, contending that RL’s generalization advantage arises from an implicit data filtering mechanism that inherently prioritizes medium-difficulty training samples. To test this hypothesis, we systematically evaluate the OOD generalization of SFT models across training datasets of varying difficulty levels. Our results confirm that data difficulty is a critical factor, revealing that training on hard samples significantly degrades OOD performance. Motivated by this finding, we introduce Difficulty-Curated SFT (DC-SFT), a straightforward method that explicitly filters the training set based on sample difficulty. Experiments show that DC-SFT not only substantially enhances OOD generalization over standard SFT, but also surpasses the performance of RL-based training, all while providing greater stability and computational efficiency. This work offers a data-centric account of the OOD generalization gap in VLMs and establishes a more efficient pathway to achieving robust generalization.

Continual Learning (CL) seeks to enable neural networks to incrementally acquire new knowledge (plasticity) while retaining existing knowledge (stability). Although pre-trained models (PTMs) have provided a strong foundation for CL, existing approaches face a fundamental challenge in balancing these two competing objectives. Current methods typically address stability by freezing the PTM backbone, which severely limits the model's plasticity, particularly when incoming data distribution diverges largely from the pre-training data. Alternatively, sequentially fine-tuning the entire PTM can adapt to new knowledge but often leads to catastrophic forgetting, highlighting the critical stability-plasticity trade-off in PTM-based CL. To address this limitation, we propose Adapting PTMs before the core CL process (ACL), a novel framework that introduces a plug-and-play adaptation phase prior to learning each new task. During this phase, ACL refines the PTM backbone by aligning embeddings with their original class prototypes while distancing them from irrelevant classes. This mechanism theoretically and empirically demonstrates desirable balance between stability and plasticity, significantly improving CL performance across benchmarks and integrated methods.

The distributed flexible job-shop scheduling problem (DFJSP) aims to coordinate job execution across distributed factories to achieve production goals. Existing reinforcement learning (RL)-based scheduling algorithms have made processes in learning adaptive scheduling polices, but rely on shallow networks and simple handcrafted rewards. These designs limit global state reasoning and accurate credit assignment under sparse rewards, thereby hindering the balanced workload distribution and efficient policy learning. To address these limitations, we propose a Large Language Model (LLM)-augmented RL algorithm tailored for DFJSP by leveraging the contextual reasoning and prior knowledge of LLM. Specifically, we propose an LLM-driven factory assignment mechanism that encodes global factory states and job features into structured queries, enabling context-aware and effective coordination among factories. Furthermore, we design an LLM-informed reward model that encodes scheduling-aware semantics into multi-dimensional proxy rewards for precise credit assignment during training. Theoretical analysis establishes bounds on the reward approximation error and demonstrates that the designed factory assignment can effectively reduce global workload variance. Moreover, extensive experiments on two benchmarks (i.e., Hurink and Brandimarte) and simulation-based DFJSP instances of varying scales demonstrate that our algorithm outperforms state-of-the-art RL algorithms, achieving the average makespan improvement ranging from 0.61% up to 25.78%.

Objectives Large language models (LLMs) have significant potential for dental applications, but their inherent tendency to hallucinate remains a major challenge. This study aims to develop and evaluate a graph-based agentic decision support system for dental question answering. Methods We developed a Dental Graph-based Agentic Decision Support System (DGADS) to support precision dental question answering. DGADS comprised three core modules: a knowledge graph builder, a graph-based RAG module, and an agentic RAG module. DGADS transformed a large volume of dental textual knowledge into a knowledge graph, termed DentalKG, and established a graph-based retrieval augmented generation system. DGADS was evaluated on three benchmarks against state-of-the-art LLMs and chunk-based RAG approaches, including internal multiple-choice questions, external multiple-choice questions, and open-ended questions. Results DentalKG comprised 130,735 entities and 236,935 triples constructed and evaluated between 15 October 2025 and 1 May 2026. DGADS leveraged knowledge from DentalKG to outperform baseline models, achieving absolute accuracy improvements of 0.04 (95% CI: 0.03, 0.06) on 500 internal questions and 0.05 (95% CI: 0.04, 0.07) on 260 external questions. In addition, DGADS improved performance on open-ended questions, with mean scores increasing up to 0.78 points on a 5-point scale. DGADS was able to automatically assess the sufficiency of information retrieved from DentalKG and retrieve relevant data from external information sources. Conclusions DGADS has the potential to become a useful research-oriented tool for supporting dental question answering, with the promise of improving the precision and efficiency of broader clinical applications. Clinical Significance DGADS supports precise dental question answering by grounding LLMs’ outputs in a dental knowledge graph and retrieving external evidence when needed. By reducing hallucinations and providing traceable rationale in benchmark-based evaluations, it may help clinicians access reliable information more efficiently in the future.

Modern manufacturers often require handling multiple tasks simultaneously under dynamic environments by sharing constrained machines. Existing multitask scheduling algorithms typically focus on transferring knowledge among multiple tasks. However, these algorithms overlook the need for collaborative multitasking efforts required to share constrained machines. To overcome this limitation, we propose a multiagent transformer (MAT)-based algorithm to solve multitask dynamic scheduling with constrained machines. Specifically, we first formulate the multitask scheduling problem as a sequential multiagent decision-making process, enabling agents to make collaborative decisions by accessing the actions of others. Furthermore, a joint policy network is developed to support the agents in adaptively selecting the appropriate heuristic for each task. It improves the decision-making quality by enabling agents to leverage common and task-specific knowledge. In addition, a comprehensive reward function is designed to guide the learning of a joint policy network for collaborative decision-making across tasks. This ensures that agents holistically consider the objectives of all tasks during the learning process. With these designs, the proposed algorithm can effectively address multiple tasks through collaborative machine sharing. The proposed algorithm is evaluated against 14 state-of-the-art competitors on 270 instances with varying scales. The results confirm that the proposed algorithm outperforms all competitors on each instance. In addition, the ablation study demonstrates the effectiveness of distinct reward mechanisms, revealing that the joint policy network makes more informed decisions by leveraging both individual and common knowledge.

OBJECTIVES:To evaluate the performance of large language models (LLMs) in risk of bias assessment and to examine whether prompt engineering improves their accuracy and alignment with expert reasoning. METHODS:We analysed 158 randomised controlled trials from 10 dental systematic reviews and their risk of bias assessments were reviewed and revised to serve as the reference standard. Two LLMs (DeepSeek-V3 and GPT-5) were evaluated under four prompting strategies, including direct command, command with reference, constrained output and formula-constrained output. The direct command served as the blank control group, simulating the approach commonly used by clinicians, whereas the other three groups employed different prompt engineering. The performance of LLMs across the seven domains of RoB-1 was evaluated using accuracy and agreement. The reasoning process of the LLMs was expressed in the form of syllogisms and its similarity to expert reasoning was assessed using MMD2. RESULTS:LLMs showed limited capability in risk of bias assessment under the blank control condition, with mean accuracies of 0.72 for DeepSeek-V3 and 0.65 for GPT-5. With formula-constrained prompting, the performance of both LLMs improved significantly, and the overall accuracy increased to 0.85 for both DeepSeek-V3 and GPT-5 (both vs the blank control group, p<0.001). Agreement metrics showed a similar pattern, with higher agreement under formula-constrained prompting than under the other prompting strategies (p<0.001 for both models). In addition, the syllogistic output format provided a clear representation of the reasoning process underlying risk of bias assessment. Compared with constrained output, formula-constrained prompting also produced reasoning that was more closely aligned with the reference answers, as indicated by lower MMD² values (DeepSeek-V3: 0.0765 vs 0.1239; GPT-5: 0.0548 vs 0.1068). CONCLUSION:Prompt engineering substantially improved the performance of LLMs in risk of bias assessment. Although LLMs cannot currently replace human reviewers, they may serve as efficient and transparent tools to support this process.

Performance evaluation is crucial in neural architecture search (NAS), but full training is costly and slow. Performance predictors offer an efficient way to quickly evaluate architectures, significantly speeding up the process. However, existing predictors often trade accuracy for speed or depend on complex encoders and costly pretraining, making it difficult to balance accuracy and efficiency with limited labeled data. In this paper, we propose a Bidirectional Multi-hop predictor with Structure-aware Ranking Loss (BMSR), which is designed for speedy and accurate performance prediction. During feature extraction, BMSR applies a bidirectional multi-hop graph convolution network with hop-aware attention to capture long-range and directional dependencies from architectures. Once the architecture embeddings are obtained, a progressively shrinking MLP is employed to compress them layer by layer, enhancing nonlinear modeling and improving representation quality. In the optimization stage, BMSR adopts a structure-aware ranking loss that leverages topological and operational similarity to encourage stable rankings among architectures. Experiments across multiple NAS benchmarks demonstrate that BMSR achieves competitive performance in both efficiency and accuracy. On NAS-Bench-201, BMSR identifies the optimal architecture using only 100 labeled samples and 8.45 s–just 0.3 https://github.com/Thorn-222/BMSR .