Unified Multimodal understanding and generation Models (UMMs) increasingly combine visual understanding and image generation within a single interactive workflow, making generated visual content available as later reasoning context. However, existing jailbreak evaluations mostly study text rewriting or isolated visual prompts, leaving the safety risk of narrative cross-turn visual grounding underexplored. We propose NarrativeAttack, a semantic-preserving visual narrative jailbreak framework. NarrativeAttack employs a three-act narrative structure in which the UMM’s own generator produces images for the setup (pre-event) and resolution (post-event) stages, making the full attack workflow self-contained while concealing the malicious event as a hidden climax. The attack concludes with an image-based "guessing game" that embeds the original malicious query among benign candidates, compelling the model to select and answer the most relevant one based on the established narrative context. A dynamic difficulty mechanism further enhances attack stability. Experiments show NarrativeAttack consistently surpasses prior approaches, achieving up to 88.25% ASR on Gemini-2.5-Flash. These results uncover an underdeveloped vulnerability and highlight the urgent need for safety alignment in UMMs.

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains underexplored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action–attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, We introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65% on average and improves benign task completion by approximately 10% under prompt injection attacks.

Aligning Large Language Models (LLMs) with human values often degrades their general capabilities, known as the alignment tax. Existing methods mitigate this by balancing dual objectives, which heavily rely on massive general-purpose data or auxiliary reward models.

In this paper, we argue that because safety features are inherently sparse within the output distribution, alignment requires localized modifications rather than global trade-offs. To this end, we propose SafeSteer, which performs on-policy distillation confined to safety tokens. First, we construct a safety teacher via activation steering. Based on this teacher, we develop a safety token selection algorithm. Consequently, SafeSteer restricts the reverse KL penalty to these tokens during training to preserve general capabilities.

Experimental results across diverse models show that our SafeSteer significantly outperforms existing methods on 7 safety benchmarks, with only minimal degradation on 5 general capability benchmarks. Notably, SafeSteer requires only 100 harmful samples without using any general-purpose data, less than 1% of what previous baselines used, considerably reducing alignment cost.

Reinforcement learning with verifiable rewards (RLVR) has substantially enhanced the reasoning capabilities of multimodal large language models (MLLMs). However, existing RLVR approaches typically rely on outcome-driven optimization that updates both perception and reasoning using a shared reward based solely on the final answer. This shared reward blurs credit assignment, frequently improving reasoning patterns while failing to reliably enhance the accuracy of upstream visual evidence extraction. To address this perception bottleneck, we introduce PRCO (Perception-Reasoning Co-evolution), a dual-role RLVR framework with a shared policy. PRCO consists of two cooperative roles: an Observer that generates an evidence caption tailored to the question and a Solver that predicts the final answer based on this caption. Crucially, PRCO employs role-specific reward signals: the Solver is optimized using verifiable outcome rewards on the final answer, while the Observer receives a utility reward derived from the Solver’s downstream success. Extensive experiments across eight challenging multimodal reasoning benchmarks demonstrate that PRCO yields consistent improvements across model scales by over 7 points on average accuracy compared to the base model, outperforming prior open-source RL-tuned baselines.

Remote sensing has emerged as an important domain for multimodal large language models (MLLMs), where reliable understanding requires locating small, dense, and spatially subtle visual evidence. However, our diagnostic analysis shows that current MLLMs often fail because their reasoning is not grounded to the correct visual evidence, leading to wrong-region attention or missed task-relevant objects. To address this issue, we introduce RsEvi-8B, an evidence-aware remote sensing MLLM that reasons with parseable bounding-box-grounded evidence. To enable this capability, we construct RsEvi-248K, a large-scale supervised dataset spanning VQA, visual grounding, object detection, object counting, and scene classification. The dataset provides verified bounding-box-grounded reasoning traces, where each intermediate visual claim is aligned with supporting evidence regions and filtered for both answer correctness and evidence faithfulness. To improve evidence-faithful reasoning, we further propose the evidence-aware GRPO, which uses an evidence verifier to regulate learning on generated evidence spans while preserving task-level outcome optimization. Extensive experiments across five remote-sensing task families show that RsEvi-8B improves the dataset-balanced average score by 17.15 points over Qwen3-VL-8B-Instruct, establishing explicit visual grounding as a key ingredient for reliable remote-sensing reasoning.

Large reasoning models (LRMs) achieve strong performance through extended reasoning traces, but they often exhibit overthinking behavior for low-complexity queries. Existing efforts to mitigate this issue are fundamentally limited by unstable accuracy–efficiency trade-offs and poor robustness to heterogeneous reasoning behaviors. To address these challenges, we propose a two-stage framework for stable adaptive thinking in LRMs. The framework first applies Hybrid Fine-Tuning to expose the model to both thinking and no-thinking behaviors, establishing well-conditioned initialization. It then performs adaptive reinforcement learning with Correctness-Preserving Advantage Shaping (CPAS) to avoid suppressing correct long-chain reasoning, and Length-Aware Gradient Regulation (LAGR) to stabilize optimization under severe reasoning-length heterogeneity. Extensive experiments on Qwen2.5-1.5B and 7B show consistent improvements over strong baselines, achieving up to +3.7/+3.6 accuracy points while reducing generated tokens by 40.6%/43.9%. Further analyses across varying problem difficulties and out-of-distribution tasks confirm the robustness and generalization of our approach.

Reinforcement learning has become a cornerstone technique for developing reasoning models in complex tasks, ranging from mathematical problem-solving to imaginary reasoning. The optimization of these models typically relies on policy gradient methods, whose efficacy hinges on the accurate estimation of an advantage function. However, prevailing methods typically employ static advantage estimation, a practice that leads to inefficient credit assignment by neglecting the dynamic utility of training samples over time. This limitation results in suboptimal policy updates, which in turn manifest as slower convergence rates and increased learning instability, as models fail to adapt to evolving sample utilities effectively. To address this problem, we introduce ADORA (Advantage Dynamics via Online Rollout Adaptation), a novel framework for policy optimization. ADORA dynamically adjusts the advantage function’s weighting by adaptively categorizing training data into temporarily advantageous and disadvantageous samples, based on their evolving utility during online model rollouts. This tailored data differentiation strategy allows ADORA to be seamlessly integrated into existing policy optimization algorithms without significant architectural modifications, enabling the policy to prioritize learning from more informative experiences and thereby achieve more efficient policy updates. Extensive evaluations across diverse model families and varying data scales demonstrate that ADORA is a robust and efficient framework. It significantly enhances long reasoning in both geometric and mathematical tasks, consistently achieving notable performance gains without requiring sensitive hyperparameter tuning.

Tabular prediction can benefit from in-table rows as few-shot evidence, yet existing tabular models typically perform instance-wise inference and LLM-based prompting is often brittle. Models do not consistently leverage relevant rows, and noisy context can degrade performance. To address this challenge, we propose TabSieve, a select-then-predict framework that makes evidence usage explicit and auditable. Given a table and a query row, TabSieve first selects a small set of informative rows as evidence and then predicts the missing target conditioned on the selected evidence. To enable this capability, we construct TabSieve-SFT-40K by synthesizing high-quality reasoning trajectories from 331 real tables using a strong teacher model with strict filtering. Furthermore, we introduce TAB-GRPO, a reinforcement learning recipe that jointly optimizes evidence selection and prediction correctness with separate rewards, and stabilizes mixed regression and classification training via dynamic task-advantage balancing. Experiments on a held-out benchmark of 75 classification and 52 regression tables show that TabSieve consistently improves performance across shot budgets, with average gains of 2.92% on classification and 4.56% on regression over the second-best baseline. Further analysis indicates that TabSieve concentrates more attention on the selected evidence, which improves robustness to noisy context.

Recent AI agents can flexibly invoke skills to solve complex tasks, but their long-term improvement is fundamentally constrained by a lack of systematic skill construction, accumulation, and transfer. In particular, without a unified framework for skill consolidation, agents tend to redundantly construct similar capabilities across different tasks, are unable to effectively transform experience into reusable assets, and struggle to generalize task-specific skills to novel scenarios. To address this limitation, we propose SkillPyramid, a skill consolidation framework that reuses existing skill experience for broader task generalization. Operating on a hierarchical skill topology, SkillPyramid further introduces a self-evolution mechanism that enables agents to compose, validate, and incorporate new skills during task execution. Experiments on ALFWorld, WebShop, and ScienceWorld across four backbone models show that SkillPyramid substantially increases the average reward by 38.0% and reduces execution steps by 27.7%. Overall, our method transforms a skill collection from a static resource pool into a dynamic evolution system.

Recent advancements in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiable reward signals, demonstrating strong performance in reasoning and code generation tasks. However, the potential safety risks associated with RLVR remain underexplored. This paper presents HarmRLVR, the first systematic investigation into the alignment reversibility risk of RLVR. We show that safety alignment can be rapidly reversed using GRPO with merely 64 harmful prompts without responses, causing models to readily comply with harmful instructions. Across five models from Llama, Qwen, and DeepSeek, we empirically demonstrate that RLVR-based attacks elevate the average harmfulness score to 4.94 with an attack success rate of 96.01%, significantly outperforming harmful fine-tuning while preserving general capabilities. Our findings reveal that RLVR can be efficiently exploited for harmful alignment, posing serious threats to open-source model safety.

While guided decoding, especially value-guided methods, has emerged as a cost-effective alternative for controlling language model outputs without re-training models, its effectiveness is limited by the accuracy of the value function. We identify that this inaccuracy stems from a core distributional gap: existing methods train static value functions on trajectories sampled exclusively from the base policy, which inherently confines their training to a narrow and suboptimal view of the potential output space. We propose Iterative Value Refinement, a novel framework designed to bridge this gap. It employs Value Exploration to provide a more comprehensive and robust training signal, complemented by Iterative Self-Refinement, which uses the improved value function from one iteration to guide the generation of higher-quality data for the next. Extensive experiments on text summarization, multi-turn dialogue, and instruction following demonstrate the effectiveness of our framework in aligning language models. Our approach not only achieves alignment but also significantly reduces computational costs by leveraging principled value function optimization for efficient and effective control.

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks. With the paradigm shift toward self-evolving agentic learning, models are increasingly expected to learn from trajectories by synthesizing tools or accumulating explicit experiences. However, prevailing methods typically rely on large-scale LLMs or multi-agent frameworks, which hinder their deployment in resource-constrained environments. The inherent sparsity of outcome-based rewards also poses a substantial challenge, as agents typically receive feedback only upon task completion. To address these limitations, we introduce a Tool-Memory based self-evolving agentic framework SEARL. Unlike approaches that directly utilize interaction experiences, our method constructs a structured experience memory that integrates planning with execution. This provides a novel form of state abstraction that facilitates the aggregation of actions within functionally analogous contexts, such as tool reuse. Consequently, agents not only extract explicit knowledge from historical data but also leverage inter-trajectory correlations to densify reward signals. We evaluate our framework on knowledge reasoning and complex search tasks, demonstrating its effectiveness in achieving more practical and efficient agentic learning.

Multimodal large language models (MLLMs) are rapidly evolving, presenting increasingly complex safety challenges. However, current dataset construction methods, which are risk-oriented, fail to cover the growing complexity of real-world multimodal safety scenarios (RMS). And due to the lack of a unified evaluation metric, their overall effectiveness remains unproven. This paper introduces a novel image-oriented self-adaptive dataset construction method for RMS, which starts with images and end constructing paired text and guidance responses. Using the image-oriented method, we automatically generate an RMS dataset comprising 35,610 image–text pairs with guidance responses. Additionally, we introduce a standardized safety dataset evaluation metric: fine-tuning a safety judge model and evaluating its capabilities on other safety datasets. Extensive experiments on various tasks demonstrate the effectiveness of the proposed image-oriented pipeline. The results confirm the scalability and effectiveness of the image-oriented approach, offering a new perspective for the construction of real-world multimodal safety datasets.

With the emergence of strong vision language capabilities, multimodal large language models (MLLMs) have demonstrated tremendous potential for real-world applications. However, the security vulnerabilities exhibited by the visual modality pose significant challenges to deploying such models in open-world environments.Recent studies have successfully induced harmful responses from target MLLMs by encoding harmful textual semantics directly into visual inputs. However, in these approaches, the visual modality primarily serves as a trigger for unsafe behavior, often exhibiting semantic ambiguity and lacking grounding in realistic scenarios. In this work, we define a novel setting: vision-centric jailbreak, where visual information serves as a necessary component in constructing a complete and realistic jailbreak context. Building on this setting, we propose the VisCo (Visual Contextual) Attack.VisCo fabricates contextual dialogue using four distinct vision-focused strategies, dynamically generating auxiliary images when necessary to construct a vision-centric jailbreak scenario.To maximize attack effectiveness, it incorporates automatic toxicity obfuscation and semantic refinement to produce a final attack prompt that reliably triggers harmful responses from the target black-box MLLMs. Specifically, VisCo achieves a toxicity score of 4.78 and an Attack Success Rate (ASR) of 85% on MM-SafetyBench against GPT-4o, significantly outperforming the baseline, which achieves a toxicity score of 2.48 and an ASR of 22.2%. Code: https://github.com/Dtc7w3PQ/Visco-Attack.

With rapid advancement and increasing accessibility of LLMs, fine-tuning aligned models has become a critical step for adapting them to real-world applications, which makes the safety of this fine-tuning process more important than ever. However, recent studies have highlighted a critical challenge: even when fine-tuning with seemingly benign downstream datasets, the safety of aligned LLMs can be compromised, making them more susceptible to malicious instructions. In this paper, we show that fine-tuning datasets often contain samples with safety-degrading features that are not easily identifiable on the surface. These samples can significantly degrade the safety alignment of LLMs during fine-tuning. To address this issue, we propose LARF, a Layer-Aware Representation Filtering method. This method identifies safety-sensitive layers within the LLM and leverages their representations to detect which data samples in the post-training dataset contain safety-degrading features. Experimental results demonstrate that LARF can effectively identify benign data with safety-degrading features. After removing such data, the safety alignment degradation caused by fine-tuning is mitigated.

In the rapidly evolving landscape of Large Language Models (LLMs), ensuring robust safety measures is paramount. To meet this crucial need, we propose SALAD-Bench, a safety benchmark specifically designed for evaluating LLMs, attack, and defense methods. Distinguished by its breadth, SALAD-Bench transcends conventional benchmarks through its large scale, rich diversity, intricate taxonomy spanning three levels, and versatile functionalities.SALAD-Bench is crafted with a meticulous array of questions, from standard queries to complex ones enriched with attack, defense modifications and multiple-choice. To effectively manage the inherent complexity, we introduce an innovative evaluators: the LLM-based MD-Judge for QA pairs with a particular focus on attack-enhanced queries, ensuring a seamless, and reliable evaluation. Above components extend SALAD-Bench from standard LLM safety evaluation to both LLM attack and defense methods evaluation, ensuring the joint-purpose utility. Our extensive experiments shed light on the resilience of LLMs against emerging threats and the efficacy of contemporary defense tactics. Data and evaluator are released under https://github.com/OpenSafetyLab/SALAD-BENCH

Multi-agent systems, when enhanced with Large Language Models (LLMs), exhibit profound capabilities in collective intelligence. However, the potential misuse of this intelligence for malicious purposes presents significant risks. To date, comprehensive research on the safety issues associated with multi-agent systems remains limited. In this paper, we explore these concerns through the innovative lens of agent psychology, revealing that the dark psychological states of agents constitute a significant threat to safety.To tackle these concerns, we propose a comprehensive framework (PsySafe) grounded in agent psychology, focusing on three key areas: firstly, identifying how dark personality traits in agents can lead to risky behaviors; secondly, evaluating the safety of multi-agent systems from the psychological and behavioral perspectives, and thirdly, devising effective strategies to mitigate these risks.Our experiments reveal several intriguing phenomena, such as the collective dangerous behaviors among agents, agents’ self-reflection when engaging in dangerous behavior, and the correlation between agents’ psychological assessments and dangerous behaviors. We anticipate that our framework and observations will provide valuable insights for further research into the safety of multi-agent systems. We make our data and code publicly accessible at https://github.com/AI4Good24/PsySafe.