As LLM agents increasingly select tools autonomously, their choices among tools with different privileges become safety-relevant. However, prior tool-selection studies focus on safety-agnostic metadata preferences, leaving privilege-sensitive choices underexplored. To address this gap, we study over-privileged tool selection, in which an agent selects or escalates to a higher-privilege tool despite a sufficient lower-privilege alternative. We introduce ToolPrivBench to evaluate whether agents choose higher-privilege tools despite sufficient lower-privilege alternatives, measuring both initial selection and escalation after transient tool failures. Across eight domains and five recurring risk patterns, we find that over-privileged tool selection is common among mainstream LLM agents and is further amplified by transient failures. We further find that general safety alignment does not reliably transfer to least-privilege tool choice, while prompt-level controls provide only limited mitigation under transient failures. We therefore introduce a privilege-aware post-training defense that teaches agents to prefer sufficient lower-privilege tools and escalate only when necessary. Our mitigation experiments show that this defense substantially reduces unnecessary high-privilege tool use while preserving general capabilities.

Reasoning models have attracted increasing attention for their ability to tackle complex tasks, embodying the System II (slow thinking) paradigm in contrast to System I (fast, intuitive responses). Yet a key question remains: Does slower reasoning necessarily lead to more truthful answers? Our findings suggest otherwise. We conduct the first systematic study of the inverse scaling law in slow-thinking paradigms for multimodal reasoning. We find that when confronted with incomplete or misleading visual inputs, slow-thinking models are more prone to fabricating plausible yet false details to justify untruthful reasoning. To analyze this behavior, we construct a 5,000-sample hierarchical prompt dataset annotated by 50 human participants. The prompts progressively increase in complexity, revealing a consistent pattern: slower reasoning models tend to follow depth-first search (DFS) thinking, persistently exploring flawed premises, while faster chat models favor breadth-first search (BFS) inference, showing greater caution under uncertainty. These findings reveal a critical vulnerability of reasoning models: while effective in structured domains such as math, their DFS-style reasoning becomes fragile when confronted with ambiguous, multimodal inputs.

Personalizing large language models often requires adapting them to new domains, tasks, or user preferences. However, such adaptation can unintentionally shift a model’s value profile, even when the training data is primarily knowledge-oriented. We refer to this failure mode as value drift, a form of value alignment tax that emerges when learning new knowledge interferes with pre-existing alignment. To address this problem, we propose VISA, a modular framework for Value Injection via Shielded Adaptation. VISA keeps the base model as a stable content generator and trains a lightweight Value Rewriter to adjust the value expression of generated responses toward a target value vector. The target is constructed through a Value Detector and an Instruction Translator, while the rewriter is optimized with Group Relative Policy Optimization using rewards for value alignment and semantic consistency. We further construct VCR-45K, a value-conditioned rewriting dataset for evaluating controllable value adaptation. Experiments show that VISA substantially improves semantic preservation over prompting and fine-tuning baselines while maintaining competitive value alignment, enabling low-cost personalized alignment with reduced value drift.

As Large Language Models (LLMs) expand in capability and application scope, their trustworthiness becomes critical. A vital risk is intrinsic deception, wherein models strategically mislead users to achieve their own objectives. Existing alignment approaches based on chain-of-thought (CoT) monitoring supervise explicit reasoning traces. However, under optimization pressure, models are incentivized to conceal deceptive reasoning, rendering semantic supervision fundamentally unreliable. Grounded in cognitive psychology, we hypothesize that a deceptive LLM maintains a stable internal belief in its CoT while its external response remains fragile under perturbation. We term this phenomenon stability asymmetry and quantify it by measuring the contrast between internal CoT stability and external response stability under perturbation. Building on this structural signature, we propose the Stability Asymmetry Regularization (SAR), a novel alignment objective that penalizes this distributional asymmetry during reinforcement learning. Unlike CoT monitoring, SAR targets the statistical structure of model outputs, rendering it robust to semantic concealment. Extensive experiments confirm that stability asymmetry reliably identifies deceptive behavior, and that SAR effectively suppresses intrinsic deception without degrading general model capability.

The rise of Omni-modality Large Language Models (OLLMs) capable of jointly processing text, audio, and visual inputs marks a major step toward general intelligence. Ensuring their alignment with human preferences requires effective Omni-modality Reward Models (ORMs), which serve as surrogates for human judgment to guide OLLMs behavior. However, ORMs evaluation remains underdeveloped in the previous literature. Existing benchmarks are largely text-centric or limited to bimodal tasks, restricting comprehensive assessment for ORMs. To bridge this gap, we introduce Omni-RewardBench, the first benchmark for comprehensive evaluation of ORMs across modalities. In short, our contributions are threefold: (1) a hybrid automatic-annotation and human-verification pipeline to construct high-quality evaluation data; (2) extensive experiments on 20+ models, including inherently omni-modal and modality-bridged systems. Our experimental results demonstrate that current OLLMs fall short as reward models, revealing several common failure modes such as perception failure, modality dominance failure, and cross-modal fusion failure; and (3) strong correlations between Omni-RewardBench scores and downstream performance (IID r = 0.94, OOD r = 0.72), validating its reliability as a predictor of real-world capability and alignment quality.

As Large Language Model (LLM) agents increasingly leverage the Model Context Protocol (MCP) to operate in complex environments, the expansion of their action spaces offers agents unsafe capabilities and underscores the risk of power-seeking. While broad action space and greater environment influence are essential for task fulfillment, they creates a fragile risk surface where minor errors or hallucinations are magnified into catastrophic failures. In response, we propose SafeMCP, a server-side defense plugin that constrains tool acquisition via predictive reasoning regarding future safety risks. SafeMCP utilizes an internal world model for look-ahead reasoning to implement a two-tier defense: proactive tool filtering to constrain hazardous power expansion and immediate intervention as a fail-safe. To train SafeMCP, we introduce a three-stage pipeline comprising environmental dynamic grounding, safe policy initialization, and reinforcement learning (RL) with dual verifiable rewards. Experiments on PowerSeeking Bench, ToolEmu, and AgentHarm show that SafeMCP achieves a safe equilibrium, effectively mitigating risks while preserving agent utility.

Large language models (LLMs) are shaping global values, yet they frequently exhibit a pronounced WEIRD (Western, Educated, Industrialized, Rich, Democratic) cultural bias, marginalizing diverse viewpoints and posing challenges for reconciling diverse populations with varying cultural backgrounds and value systems. In this work, we move beyond simple alignment methods to propose a new paradigm for cross-cultural fairness. We introduce a Nash Consensus Negotiation framework under the formulation of cross-cultural consensus as a Nash Equilibrium. Each LLM iteratively proposes and refines natural-language guidelines, guided by a utility function balancing self-consistency with mutual acceptance, while penalizing redundancy. The process expands the proposal space and converges to a consensus, yielding fair and interpretable consensus outcomes. We evaluate our framework against baselines using quantitative metrics, qualitative analysis, and large-scale human studies. Experiments demonstrate that our framework generates higher-quality and more balanced consensus, effectively mitigating assimilation toward WEIRD values. Furthermore, we finetune diverse LLM architectures with negotiation data via preference optimization and supervised reasoning, reducing cultural distances by up to 95.53%. Overall, our work offers a systematic path to mitigate cultural bias in LLMs by guiding them toward self-consistency, mutually-acceptable equilibria.

The advent of multi-modal large language models (MLLMs) has greatly advanced research on video fake news detection (VFND) tasks. Existing benchmarks typically focus on the detection accuracy, while failing to provide fine-grained assessments for the entire detection process. To address these limitations, we introduce POVFNDB (Process-oriented Video Fake News Detection Benchmark), a process-oriented benchmark comprising 10 tasks designed to systematically evaluate MLLMs’ perception, understanding, and reasoning capabilities in VFND. This benchmark contains 36,240 human-annotated question-answer (QA) in structured or open-ended formats, spanning 15 distinct evaluation dimensions that characterize different aspects of the video fake news detection process.Using POVFNDB, we conduct comprehensive evaluations on both proprietary and open-source MLLMs. Moreover, We fine-tune Qwen2.5VL-7B-Instruct on a reasoning dataset generated by our proposed POVFND-CoT, a chain-of-thought method that utilizes rationales from evaluation results and rationale validation. The resulting model achieves sota performance on VFND.

Multimodal Process Reward Models (MPRMs) have emerged as a pivotal framework for enhancing the reasoning capabilities of Multimodal Large Language Models (MLLMs). However, the research community currently lacks a dedicated benchmark to rigorously assess the error discernment capabilities of these models.To address this gap, we introduce PRMBench-V, a novel benchmark specifically designed to evaluate MPRMs’ proficiency in detecting erroneous reasoning steps across diverse error categories. Leveraging a semi-automated annotation pipeline augmented with human verification, we construct a comprehensive dataset comprising 907 unique queries, each annotated with nine distinct error types, resulting in 8,163 test cases with fine-grained step-level error labels.Through extensive experiments involving over 15 open- and closed-source models, we uncover several key findings: (1) even the strongest existing MPRMs achieve only \textasciitilde30% accuracy in error identification; (2) while partial error detection achieves moderate precision and recall (\textasciitilde60%), overall accuracy remains low (\textasciitilde20%); and (3) benchmark scores exhibit a strong correlation with downstream task performance gains (r=0.86). Furthermore, we demonstrate that PRMBench-V can inform the development of more robust MPRMs: by introducing the Bayesian Rater Reliability Process Reward Model (BR2-PRM), we achieve up to a 4.8% performance improvement through test-time scaling.We believe that PRMBench-V will serve as a valuable resource for advancing MPRM research, enabling more rigorous evaluation and fostering the development of models with fine-grained multimodal reasoning capabilities.

With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn Attack Success Rate (ASR) compared to existed guard models.

We present a novel three-stage framework leveraging Large Language Models (LLMs) within a risk-aware multi-agent system for automate strategy finding in quantitative finance. Our approach addresses the brittleness of traditional deep learning models in financial applications by: employing prompt-engineered LLMs to generate executable alpha factor candidates across diverse financial data, implementing multimodal agent-based evaluation that filters factors based on market status, predictive quality while maintaining category balance, and deploying dynamic weight optimization that adapts to market conditions. Experimental results demonstrate the robust performance of the strategy in Chinese & US market regimes compared to established benchmarks. Our work extends LLMs capabilities to quantitative trading, providing a scalable architecture for financial signal extraction and portfolio construction. The overall framework significantly outperforms all benchmarks with 53.17% cumulative return on SSE50 (Jan 2023 to Jan 2024), demonstrating superior risk-adjusted performance and downside protection on the market.

Existing alignment methods share a common topology of information flow, where reward information is collected from humans, modeled with preference learning, and used to tune language models. However, this shared topology has not been systematically characterized, nor have its alternatives been thoroughly explored, leaving the problems of low data efficiency and unreliable generalization unaddressed. As a solution, we introduce a theory of reward generalization in reinforcement learning from human feedback (RLHF), focusing on the topology of information flow at both macro and micro levels. At the macro level, we portray the RLHF information flow as an autoencoding process over behavior distributions, formalizing the RLHF objective of distributional consistency between human preference and model behavior. At the micro level, we present induced Bayesian networks to model the impact of dataset topologies on reward generalization. Combining analysis on both levels, we propose reward modeling from tree-structured preference information. It is shown to reduce reward uncertainty by up to 𝛩(log n/loglog n) times compared to baselines, where n is the dataset size. Validation on three NLP tasks shows that it achieves an average win rate of 65% against baselines, thus improving reward generalization for free via topology design, while reducing the amount of data requiring annotation.

With the growing prevalence of large language models (LLMs), the safety of LLMs has raised significant concerns. However, there is still a lack of definitive standards for evaluating their safety due to the subjective nature of current safety benchmarks. To address this gap, we conducted the first exploration of LLMs’ safety evaluation from a legal perspective by proposing the SafeLawBench benchmark. SafeLawBench categorizes safety risks into three levels based on legal standards, providing a systematic and comprehensive framework for evaluation. It comprises 24,860 multi-choice questions and 1,106 open-domain question-answering (QA) tasks. Our evaluation included 2 closed-source LLMs and 18 open-source LLMs using zero-shot and few-shot prompting, highlighting the safety features of each model. We also evaluated the LLMs’ safety-related reasoning stability and refusal behavior. Additionally, we found that a majority voting mechanism can enhance model performance. Notably, even leading SOTA models like Claude-3.5-Sonnet and GPT-4o have not exceeded 80.5% accuracy in multi-choice tasks on SafeLawBench, while the average accuracy of 20 LLMs remains at 68.8%. We urge the community to prioritize research on the safety of LLMs.

Large language models (LLMs) have shown promising first-order logic (FOL) reasoning capabilities with applications in various areas. However, their effectiveness in complex mathematical reasoning involving multi-step FOL deductions is still under-researched. While LLMs perform competitively on established mathematical reasoning benchmarks, they struggle with multi-step FOL tasks, as demonstrated by Deepseek-Prover-V2-7B’s low accuracy (4.2%) on our proposed theorem proving dataset. This issue arises from the limited exploration of diverse proof strategies and the potential for early reasoning mistakes to undermine entire proofs. To address these issues, we propose DREAM, a self-adaptive solution that enhances the Diversity and REAsonability of LLMs’ generation strategies. DREAM incorporates an Axiom-Driven Strategy Diversification mechanism to promote varied strategic outcomes and a Sub-Proposition Error Feedback to help LLMs reflect on and correct their proofs. Our contributions include pioneering advancements in LLMs’ mathematical reasoning through FOL theorem proving, introducing a novel inference stage solution that improves performance by 0.6% to 6.4%, and providing a curated dataset of 447 mathematical theorems in Lean 4 format for evaluation.

In this work, we introduce the PKU-SafeRLHF dataset, designed to promote research on safety alignment in large language models (LLMs). As a sibling project to SafeRLHF and BeaverTails, we separate annotations of helpfulness and harmlessness for question-answering pairs, providing distinct perspectives on these coupled attributes. Overall, we provide 44.6k refined prompts and 265k question-answer pairs with safety meta-labels for 19 harm categories and three severity levels ranging from minor to severe, with answers generated by Llama-family models. Based on this, we collected 166.8k preference data, including dual-preference (helpfulness and harmlessness decoupled) and single-preference data (trade-off the helpfulness and harmlessness from scratch), respectively. Using the large-scale annotation data, we further train severity-sensitive moderation for the risk control of LLMs and safety-centric RLHF algorithms for the safety alignment of LLMs. We believe this dataset will be a valuable resource for the community, aiding in the safe deployment of LLMs.

Large language models (LLMs) may exhibit unintended or undesirable behaviors. Recent works have concentrated on aligning LLMs to mitigate harmful outputs. Despite these efforts, some anomalies indicate that even a well-conducted alignment process can be easily circumvented, whether intentionally or accidentally. Does alignment fine-tuning yield have robust effects on models, or are its impacts merely superficial? In this work, we make the first exploration of this phenomenon from both theoretical and empirical perspectives. Empirically, we demonstrate the elasticity of post-alignment models, i.e., the tendency to revert to the behavior distribution formed during the pre-training phase upon further fine-tuning. Leveraging compression theory, we formally deduce that fine-tuning disproportionately undermines alignment relative to pre-training, potentially by orders of magnitude. We validate the presence of elasticity through experiments on models of varying types and scales. Specifically, we find that model performance declines rapidly before reverting to the pre-training distribution, after which the rate of decline drops significantly. Furthermore, we further reveal that elasticity positively correlates with the increased model size and the expansion of pre-training data. Our findings underscore the need to address the inherent elasticity of LLMs to mitigate their resistance to alignment.