While Reinforcement Learning from Human Feedback (RLHF) has become the predominant method for controlling language model outputs, it suffers from high computational costs and training instability. Guided decoding, especially value-guided methods, offers a cost-effective alternative by controlling outputs without re-training models. However, the accuracy of the value function is crucial for value-guided decoding, as inaccuracies can lead to suboptimal decision-making and degraded performance. Existing methods struggle with accurately estimating the optimal value function, leading to less effective control. We propose Iterative Value Function Optimization, a novel framework that addresses these limitations through two key components: Monte Carlo Value Estimation, which reduces estimation variance by exploring diverse trajectories, and Iterative On-Policy Optimization, which progressively improves value estimation through collecting trajectories from value-guided policies. Extensive experiments on text summarization, multi-turn dialogue, and instruction following demonstrate the effectiveness of value-guided decoding approaches in aligning language models. These approaches not only achieve alignment but also significantly reduce computational costs by leveraging principled value function optimization for efficient and effective control.

LLM-based multi-agent systems (MAS) have shown significant potential in tackling diverse tasks. However, to design effective MAS, existing approaches heavily rely on manual configurations or multiple calls of advanced LLMs, resulting in inadaptability and high inference costs. In this paper, we simplify the process of building an MAS by reframing it as a generative language task, where the input is a user query and the output is a corresponding MAS. To address this novel task, we unify the representation of MAS as executable code and propose a consistency-oriented data construction pipeline to create a high-quality dataset comprising coherent and consistent query-MAS pairs. Using this dataset, we train MAS-GPT, an open-source medium-sized LLM that is capable of generating query-adaptive MAS within a single LLM inference. The generated MAS can be seamlessly applied to process user queries and deliver high-quality responses. Extensive experiments on 9 benchmarks and 5 LLMs show that the proposed MAS-GPT consistently outperforms 10+ baseline MAS methods on diverse settings, indicating MAS-GPT's high effectiveness, efficiency and strong generalization ability. Code will be available at https://github.com/rui-ye/MAS-GPT.

Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: \textbf{neuron misidentification} due to simplistic parameter magnitude-based selection, and \textbf{cross-task neuron interference} during merging. To address these challenges, we propose \textbf{LED-Merging}, a three-stage framework that \textbf{L}ocates task-specific neurons via gradient-based attribution, dynamically \textbf{E}lects critical neurons through multi-model importance fusion, and \textbf{D}isjoints conflicting updates through parameter isolation. Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging reduces harmful response rates(\emph{e.g.}, a 31.4\% decrease on Llama-3-8B-Instruct on HarmBench) while preserving 95\% of utility performance(\emph{e.g.}, 52.39\% accuracy on GSM8K). LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs.

Despite the rapid development of safety alignment techniques for LLMs, defending against multi-turn jailbreaks is still a challenging task. In this paper, we conduct a comprehensive comparison, revealing that some existing defense methods can improve the robustness of LLMs against multi-turn jailbreaks but compromise usability, i.e., reducing general capabilities or causing the over-refusal problem. From the perspective of mechanism interpretability of LLMs, we discover that these methods fail to establish a boundary that exactly distinguishes safe and harmful feature representations. Therefore, boundary-safe representations close to harmful representations are inevitably disrupted, leading to a decline in usability. To address this issue, we propose X-Boundary to push harmful representations away from boundary-safe representations and obtain an exact distinction boundary. In this way, harmful representations can be precisely erased without disrupting safe ones. Experimental results show that X-Boundary achieves state-of-the-art defense performance against multi-turn jailbreaks, while reducing the over-refusal rate by about 20% and maintaining nearly complete general capability. Furthermore, we theoretically prove and empirically verify that X-Boundary can accelerate the convergence process during training. Please see our code at: https://github.com/AI45Lab/X-Boundary.

Explaining the hidden representations of Large Language Models (LLMs) is a perspective to understand LLMs' underlying inference logic and improve their reliability in application scenarios. However, previous methods introduce external ''black-box'' modules to explain ''black-box'' LLMs, increasing the potential uncertainty and failing to provide faithful explanations. In this paper, we propose a self-explaining method SEER, enhancing LLMs' explainability by aggregating the same concept and disentangling the different concepts in the representation space. In this way, SEER provides faithful explanations carried by representations synchronously with the LLMs' output. Additionally, we showcase the applications of SEER on trustworthiness-related tasks (e.g., the safety risks classification and detoxification tasks), where self-explained LLMs achieve consistent improvement in explainability and performance. More crucially, we theoretically analyze the improvement of SEER on LLMs' generalization ability through optimal transport theory.

Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin. Data and Models are released under: \href{https://dripnowhy.github.io/MIS/}{\texttt{https://dripnowhy.github.io/MIS/}}

Large Language Models (LLMs) demonstrate remarkable zero-shot performance across various natural language processing tasks. The integration of multimodal encoders extends their capabilities, enabling the development of Multimodal Large Language Models that process vision, audio, and text. However, these capabilities also raise significant security concerns, as these models can be manipulated to generate harmful or inappropriate content through jailbreak. While extensive research explores the impact of modality-specific input edits on text-based LLMs and Large Vision-Language Models in jailbreak, the effects of audio-specific edits on Large Audio-Language Models (LALMs) remain underexplored. Hence, this paper addresses this gap by investigating how audio-specific edits influence LALMs inference regarding jailbreak. We introduce the Audio Editing Toolbox (AET), which enables audio-modality edits such as tone adjustment, word emphasis, and noise injection, and the Edited Audio Datasets (EADs), a comprehensive audio jailbreak benchmark. We also conduct extensive evaluations of state-of-the-art LALMs to assess their robustness under different audio edits. This work lays the groundwork for future explorations on audio-modality interactions in LALMs security.

Text-to-image (T2I) models have rapidly advanced, enabling the generation of high-quality images from text prompts across various domains. However, these models present notable safety concerns, including the risk of generating harmful, biased, or private content. Current research on assessing T2I safety remains in its early stages. While some efforts have been made to evaluate models on specific safety dimensions, many critical risks remain unexplored. To address this gap, we introduce T2ISafety, a safety benchmark that evaluates T2I models across three key domains: toxicity, fairness, and bias. We build a detailed hierarchy of 12 tasks and 44 categories based on these three domains, and meticulously collect 70K corresponding prompts. Based on this taxonomy and prompt set, we build a large-scale T2I dataset with 68K manually annotated images and train an evaluator capable of detecting critical risks that previous work has failed to identify, including risks that even ultra-large proprietary models like GPTs cannot correctly detect. We evaluate 12 prominent diffusion models on T2ISafety and reveal several concerns including persistent issues with racial fairness, a tendency to generate toxic content, and significant variation in privacy protection across the models, even with defense methods like concept erasing. Data and evaluator are released under https://github.com/adwardlee/t2i_safety.

With the integration of large language models (LLMs), embodied agents have strong capabilities to execute complicated instructions in natural language, paving a way for the potential deployment of embodied robots. However, a foreseeable issue is that those embodied agents can also flawlessly execute some hazardous tasks, potentially causing damages in real world. To study this issue, we present SafeAgentBench -- a new benchmark for safety-aware task planning of embodied LLM agents. SafeAgentBench includes: (1) a new dataset with 750 tasks, covering 10 potential hazards and 3 task types; (2) SafeAgentEnv, a universal embodied environment with a low-level controller, supporting multi-agent execution with 17 high-level actions for 8 state-of-the-art baselines; and (3) reliable evaluation methods from both execution and semantic perspectives. Experimental results show that the best-performing baseline gets 69% success rate for safe tasks, but only 5% rejection rate for hazardous tasks, indicating significant safety risks. More details and codes are available at https://github.com/shengyin1224/SafeAgentBench.

Aligning the behaviors of Multimodal Large Language Models (MLLMs) with human preferences is crucial for developing robust and trustworthy AI systems. While recent attempts have employed human experts or powerful auxiliary AI systems to provide more accurate preference feedback, such as determining the preferable responses from MLLMs or directly rewriting hallucination-free responses, extensive resource overhead compromise the scalability of the feedback collection. In this work, we introduce Topic-level Preference Overwriting (TPO), a self-correctional approach that guide the model itself to mitigate its own hallucination at the topic level. Through a deconfounded strategy that replaces each topic within the response with the best or worst alternatives generated by the model itself, TPO creates more contrasting pairwise preference feedback, enhancing the feedback quality without human or proprietary model intervention. Notably, the experimental results demonstrate proposed TPO achieves state-of-the-art performance in trustworthiness, significantly reducing the object hallucinations by 92% and overall hallucinations by 38%. Code, model and dataset are available now.