Federated learning (FL) enables collaborative training across decentralized devices, but under non-IID data it faces a persistent tension between adversarial robustness, communication efficiency, and utility on rare classes. We present AgenticFedLock, a trust-conditioned FL framework that treats robustness as sequential control of trainable parameter-space geometry through a hierarchical server-side protocol controller. Its four coordinated agents estimate client state, select participants, allocate trainable subspaces, and evaluate or repair risky updates. The Judge-Repair agent’s backdoor-risk signal is trigger-agnostic by construction: it fuses subspace-localized root-cosine deviation, trusted-consensus deviation, and norm-inflation detection, none of which use trigger or target information, combined with a norm-confidence-gated root-anchor reweighting at aggregation. This signal is verified to trigger repair in the large majority of completed rounds under an unknown-defender setting, with no dependence on privileged knowledge of the attack. Across a matched three-seed, five-method, four-dataset, two-attack unknown-defender comparison, the controller reduces model-replacement attack success to 15.6±6.2% on EMNIST Balanced and 26.1±17.2% on CIFAR-10, where FedAvg, FLTrust, FLAME, and RFA all remain above 90%; on MNIST and Fashion-MNIST it is competitive with the strongest baselines under both attacks studied. It does so while transmitting 60%–67% of protocol bytes against 100% for every full-model comparator. Beyond suppressing the attack’s effect, the same persistent per-client state lets the controller identify which clients are malicious with near-perfect accuracy under model replacement (suspicion-score AUC 0.999–1.000, reaching near-ceiling separation within the first few rounds), a capability the compared robust-aggregation baselines do not expose since none of them persist an identity-linked judgment across rounds; this attribution collapses to near-chance under the camouflaged adaptive-mimic attack. Matched ablation, sensitivity, defender-knowledge, and systems-cost results accompany these findings, including the cases where the mechanism does not close the gap.
Recent advances in multimodal large language models (MLLMs) have demonstrated remarkable multi- and cross-modal integration capabilities. However, their potential for fine-grained emotion understanding remains systematically underexplored. While open-vocabulary multimodal emotion recognition (MER-OV) has emerged as a promising direction to overcome the limitations of closed emotion sets, no comprehensive evaluation of MLLMs in this context currently exists. To address this, our work presents the first large-scale benchmarking study of MER-OV on the OV-MERD dataset, evaluating 19 mainstream MLLMs, including general-purpose, modality-specialized, and reasoning-enhanced architectures. Through systematic analysis of model reasoning capacity, fusion strategies, contextual utilization, and prompt design, we provide key insights into the capabilities and limitations of current MLLMs for MER-OV. Our evaluation reveals that a two-stage, trimodal (audio, video, and text) fusion approach achieves optimal performance in MER-OV, with video emerging as the most critical modality. We further identify a surprisingly narrow gap between open- and closed-source LLMs. These findings establish essential benchmarks and offer practical guidelines for advancing open-vocabulary and fine-grained affective computing, paving the way for more nuanced and interpretable emotion AI systems. Associated code will be made publicly available upon acceptance.
The emergence of large language models (LLMs) has significantly advanced natural language understanding, generation, and reasoning across various domains, including the biomedical field. Despite these advancements, the evaluation of biomedical LLMs remains limited, primarily relying on manually crafted datasets which are insufficient for comprehensively assessing LLMs’ capabilities. To address this challenge and cater specifically to the requirements of biomedical LLMs, we propose KGMedQA, an innovative evaluation benchmark based on knowledge graphs (KGs) designed to assess the knowledge and reasoning abilities of LLMs. Through careful alignment between natural language and KG structures, KGMedQA can be applied to arbitrary KGs, enabling automated question/answer generation and LLM evaluation. By leveraging the advantages of KG structures, we design seven tasks of varying complexity and focus, accompanied by specialized evaluation metrics. Experiments conducted with KGMedQA involve ten different LLMs, including general and specialized biomedical models, tested across two KGs focusing on different types of biomedical knowledge and reasoning. Compared to traditional methods, our results uncover more novel insights. For instance, while specialized models exhibit strengths in knowledge, they have deficiencies in reasoning abilities compared to general models. Additionally, factors such as model scales and prompting methods also impact the performance of LLMs. Our benchmark represents advancements in the evaluation of domain-specific LLMs, offering an effective tool for future research and development. Our source code is available at: https://github.com/PerseidsMeteorShower/KGMedQA.
While rapid advances in large language models (LLMs) are reshaping data-driven intelligent education, accurately simulating students remains an important but challenging bottleneck for scalable educational data collection, evaluation, and intervention design. However, current works are limited by scarce real interaction data, costly expert evaluation for realism, and a lack of large-scale, systematic analyses of LLMs ability in simulating students. We address this gap by presenting a three-stage LLM-human collaborative pipeline to automatically generate and filter high-quality student agents. We leverage a two-round automated scoring validated by human experts and deploy a score propagation module to obtain more consistent scores across the student similarity graph. Experiments show that combining automated scoring, expert calibration, and graph-based propagation yields simulated student that more closely track authentication by human judgments. We then analyze which profiles and behaviors are simulated more faithfully, supporting subsequent studies on personalized learning and educational assessment.
Project compliance review serves as a critical component in ensuring that project submissions meet regulatory and procedural standards while filtering out non-compliant proposals. However, the rapid increase in project applications has rendered traditional manual review mechanisms inefficient and unsustainable. Existing automated approaches—based on either semantic vector matching or direct use of large language models (LLMs)—often struggle with heterogeneous document structures, limited robustness, and high computational cost. To overcome these challenges, this study proposes an automated compliance review framework based on multi-agent routing. The framework integrates multiple expert agents with distinct reasoning paradigms and employs a classification agent to dynamically route each task to the most suitable expert, thereby enabling data-adaptive decision-making. Through adaptive routing, the system effectively reduces redundant computation by invoking complex reasoning only when necessary. Experimental results on real-world datasets from the power industry demonstrate that our method achieves superior accuracy and 48% reduction in token consumption compared to the best-performing single-agent baseline, achieving a balanced trade-off between effectiveness and efficiency. Furthermore, an online review platform has been developed and successfully deployed in large-scale power project evaluation scenarios, validating the practicality and scalability of the proposed approach.
