Despite significant progress in safety alignment, large language models (LLMs) remain susceptible to jailbreak attacks. Existing defense mechanisms have not fully deleted harmful knowledge in LLMs, which allows such attacks to bypass safeguards and produce harmful outputs. To address this challenge, we propose a novel safety alignment strategy, Constrained Knowledge Unlearning (CKU), which focuses on two primary objectives: knowledge localization and retention, and unlearning harmful knowledge. CKU works by scoring neurons in specific multilayer perceptron (MLP) layers to identify a subset U of neurons associated with useful knowledge. During the unlearning process, CKU prunes the gradients of neurons in U to preserve valuable knowledge while effectively mitigating harmful content. Experimental results demonstrate that CKU significantly enhances model safety without compromising overall performance, offering a superior balance between safety and utility compared to existing methods. Additionally, our analysis of neuron knowledge sensitivity across various MLP layers provides valuable insights into the mechanics of safety alignment and model knowledge editing.

Retrieval-Augmented Generation (RAG) enhances factual accuracy by integrating external knowledge, yet it introduces a critical issue: erroneous or biased retrieval can mislead generation, compounding hallucinations, a phenomenon we term Hallucination on Hallucination. To address this, we propose Debate-Augmented RAG (DRAG), a training-free framework that integrates Multi-Agent Debate (MAD) mechanisms into both retrieval and generation stages. In retrieval, DRAG employs structured debates among proponents, opponents, and judges to refine retrieval quality and ensure factual reliability. In generation, DRAG introduces asymmetric information roles and adversarial debates, enhancing reasoning robustness and mitigating factual inconsistencies. Evaluations across multiple tasks demonstrate that DRAG improves retrieval reliability, reduces RAG-induced hallucinations, and significantly enhances overall factual accuracy. Our code is available at https://github.com/Huenao/Debate-Augmented-RAG.

Probing techniques have shown promise in revealing how LLMs encode human-interpretable concepts, particularly when applied to curated datasets. However, the factors governing a dataset's suitability for effective probe training are not well-understood. This study hypothesizes that probe performance on such datasets reflects characteristics of both the LLM's generated responses and its internal feature space. Through quantitative analysis of probe performance and LLM response uncertainty across a series of tasks, we find a strong correlation: improved probe performance consistently corresponds to a reduction in response uncertainty, and vice versa. Subsequently, we delve deeper into this correlation through the lens of feature importance analysis. Our findings indicate that high LLM response variance is associated with a larger set of important features, which poses a greater challenge for probe models and often results in diminished performance. Moreover, leveraging the insights from response uncertainty analysis, we are able to identify concrete examples where LLM representations align with human knowledge across diverse domains, offering additional evidence of interpretable reasoning in LLMs.

Large Language Models (LLMs)-based Multi-Agent Systems (MAS) exhibit remarkable problem-solving and task planning capabilities across diverse domains due to their specialized agentic roles and collaborative interactions. However, this also amplifies the severity of security risks under MAS attacks. To address this, we introduce MASTER, a novel security research framework for MAS, focusing on diverse Role configurations and Topological structures across various scenarios. MASTER offers an automated construction process for different MAS setups and an information-flow-based interaction paradigm. To tackle MAS security challenges in varied scenarios, we design a scenario-adaptive, extensible attack strategy utilizing role and topological information, which dynamically allocates targeted, domain-specific attack tasks for collaborative agent execution. Our experiments demonstrate that such an attack, leveraging role and topological information, exhibits significant destructive potential across most models. Additionally, we propose corresponding defense strategies, substantially enhancing MAS resilience across diverse scenarios. We anticipate that our framework and findings will provide valuable insights for future research into MAS security challenges.

The rapid advancement of Large Language Models (LLMs) has resulted in interest in their potential applications within manufacturing systems, particularly in the context of Industry 5.0. However, determining when to implement LLMs versus other Natural Language Processing (NLP) techniques, ontologies or knowledge graphs, remains an open question. This paper offers decision-making guidance for selecting the most suitable technique in various industrial contexts, emphasizing human-robot collaboration and resilience in manufacturing. We examine the origins and unique strengths of LLMs, ontologies, and knowledge graphs, assessing their effectiveness across different industrial scenarios based on the number of domains or disciplines required to bring a product from design to manufacture. Through this comparative framework, we explore specific use cases where LLMs could enhance robotics for human-robot collaboration, while underscoring the continued relevance of ontologies and knowledge graphs in low-dependency or resource-constrained sectors. Additionally, we address the practical challenges of deploying these technologies, such as computational cost and interpretability, providing a roadmap for manufacturers to navigate the evolving landscape of Language based AI tools in Industry 5.0. Our findings offer a foundation for informed decision-making, helping industry professionals optimize the use of Language Based models for sustainable, resilient, and human-centric manufacturing. We also propose a Large Knowledge Language Model architecture that offers the potential for transparency and configuration based on complexity of task and computing resources available.

We present MSA-MathEval, our submission to the BEA 2025 Shared Task on evaluating AI tutor responses across four instructional dimensions: Mistake Identification, Mistake Location, Providing Guidance, and Actionability. Our approach uses a unified training pipeline to fine-tune a single instruction-tuned language model across all tracks, without any task-specific architectural changes. To improve prediction reliability, we introduce a disagreement-aware ensemble inference strategy that enhances coverage of minority labels. Our system achieves strong performance across all tracks, ranking 1st in Providing Guidance, 3rd in Actionability, and 4th in both Mistake Identification and Mistake Location. These results demonstrate the effectiveness of scalable instruction tuning and disagreement-driven modeling for robust, multi-dimensional evaluation of LLMs as educational tutors.

Recent advances in cross-prompt automated essay scoring (AES) typically train models jointly on all source prompts, often requiring additional access to unlabeled target prompt essays simultaneously. However, using all sources is suboptimal in our pilot study, and re-accessing source datasets during adaptation raises privacy concerns. We propose a source-free adaptation approach that selectively merges individually trained source models' parameters instead of datasets. In particular, we simulate joint training through linear combinations of task vectors -- the parameter updates from fine-tuning. To optimize the combination's coefficients, we propose Prior-encoded Information Maximization (PIM), an unsupervised objective which promotes the model's score discriminability regularized by priors pre-computed from the sources. We employ Bayesian optimization as an efficient optimizer of PIM. Experimental results with LLMs on in-dataset and cross-dataset adaptation show that our method (1) consistently outperforms training jointly on all sources, (2) maintains superior robustness compared to other merging methods, (3) excels under severe distribution shifts where recent leading cross-prompt methods struggle, all while retaining computational efficiency.

Large language models (LLMs) often exhibit strong biases, e.g, against women or in favor of the number 7. We investigate whether LLMs would be able to output less biased answers when allowed to observe their prior answers to the same question in a multi-turn conversation. To understand which types of questions invite more biased answers, we test LLMs on our proposed set of questions that span 9 topics and belong to three types: (1) Subjective; (2) Random; and (3) Objective. Interestingly, LLMs are able to "de-bias" themselves in a multi-turn conversation in response to questions that seek an Random, unbiased answer. Furthermore, we propose B-score, a novel metric that is effective in detecting biases to Subjective, Random, Easy, and Hard questions. On MMLU, HLE, and CSQA, leveraging B-score substantially improves the verification accuracy of LLM answers (i.e, accepting LLM correct answers and rejecting incorrect ones) compared to using verbalized confidence scores or the frequency of single-turn answers alone. Code and data are available at: https://b-score.github.io.

Retrieval-Augmented Generation (RAG) has proven effective in mitigating hallucinations in large language models by incorporating external knowledge during inference. However, this integration introduces new security vulnerabilities, particularly to poisoning attacks. Although prior work has explored various poisoning strategies, a thorough assessment of their practical threat to RAG systems remains missing. To address this gap, we propose the first comprehensive benchmark framework for evaluating poisoning attacks on RAG. Our benchmark covers 5 standard question answering (QA) datasets and 10 expanded variants, along with 13 poisoning attack methods and 7 defense mechanisms, representing a broad spectrum of existing techniques. Using this benchmark, we conduct a comprehensive evaluation of all included attacks and defenses across the full dataset spectrum. Our findings show that while existing attacks perform well on standard QA datasets, their effectiveness drops significantly on the expanded versions. Moreover, our results demonstrate that various advanced RAG architectures, such as sequential, branching, conditional, and loop RAG, as well as multi-turn conversational RAG, multimodal RAG systems, and RAG-based LLM agent systems, remain susceptible to poisoning attacks. Notably, current defense techniques fail to provide robust protection, underscoring the pressing need for more resilient and generalizable defense strategies.

Large Language Models (LLMs) have shown promise in character imitation, enabling immersive and engaging conversations. However, they often generate content that is irrelevant or inconsistent with a character's background. We attribute these failures to: (1) the inability to accurately recall character-specific knowledge due to entity ambiguity, and (2) a lack of awareness of the character's cognitive boundaries. To address these issues, we propose RoleRAG, a retrieval-based framework that integrates efficient entity disambiguation for knowledge indexing with a boundary-aware retriever for extracting contextually appropriate information from a structured knowledge graph. Experiments on role-playing benchmarks show that RoleRAG's calibrated retrieval helps both general-purpose and role-specific LLMs better align with character knowledge and reduce hallucinated responses.