Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards

Large language models for mathematical reasoning are typically trained with outcome-based rewards, which credit only the final answer. In our experiments, we observe that this paradigm is highly susceptible to reward hacking, leading to a substantial overestimation of a model's reasoning ability. This is evidenced by a high incidence of false positives - solutions that reach the correct final answer through an unsound reasoning process. Through a systematic analysis with human verification, we establish a taxonomy of these failure modes, identifying patterns like Miracle Steps - abrupt jumps to a correct output without a valid preceding derivation. Probing experiments suggest a strong association between these Miracle Steps and memorization, where the model appears to recall the answer directly rather than deriving it. To mitigate this systemic issue, we introduce the Rubric Reward Model (RRM), a process-oriented reward function that evaluates the entire reasoning trajectory against problem-specific rubrics. The generative RRM provides fine-grained, calibrated rewards (0-1) that explicitly penalize logical flaws and encourage rigorous deduction. When integrated into a reinforcement learning pipeline, RRM-based training consistently outperforms outcome-only supervision across four math benchmarks. Notably, it boosts Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. Our work demonstrates that rewarding the solution process is crucial for building models that are not only more accurate but also more reliable.

Towards Evaluating Proactive Risk Awareness of Multimodal Language Models

Human safety awareness gaps often prevent the timely recognition of everyday risks. In solving this problem, a proactive safety artificial intelligence (AI) system would work better than a reactive one. Instead of just reacting to users' questions, it would actively watch people's behavior and their environment to detect potential dangers in advance. Our Proactive Safety Bench (PaSBench) evaluates this capability through 416 multimodal scenarios (128 image sequences, 288 text logs) spanning 5 safety-critical domains. Evaluation of 36 advanced models reveals fundamental limitations: Top performers like Gemini-2.5-pro achieve 71% image and 64% text accuracy, but miss 45-55% risks in repeated trials. Through failure analysis, we identify unstable proactive reasoning rather than knowledge deficits as the primary limitation. This work establishes (1) a proactive safety benchmark, (2) systematic evidence of model limitations, and (3) critical directions for developing reliable protective AI. We believe our dataset and findings can promote the development of safer AI assistants that actively prevent harm rather than merely respond to requests. Our dataset can be found at https://huggingface.co/datasets/Youliang/PaSBench.

VisBias: Measuring Explicit and Implicit Social Biases in Vision Language Models

This research investigates both explicit and implicit social biases exhibited by Vision-Language Models (VLMs). The key distinction between these bias types lies in the level of awareness: explicit bias refers to conscious, intentional biases, while implicit bias operates subconsciously. To analyze explicit bias, we directly pose questions to VLMs related to gender and racial differences: (1) Multiple-choice questions based on a given image (e.g., "What is the education level of the person in the image?") (2) Yes-No comparisons using two images (e.g., "Is the person in the first image more educated than the person in the second image?") For implicit bias, we design tasks where VLMs assist users but reveal biases through their responses: (1) Image description tasks: Models are asked to describe individuals in images, and we analyze disparities in textual cues across demographic groups. (2) Form completion tasks: Models draft a personal information collection form with 20 attributes, and we examine correlations among selected attributes for potential biases. We evaluate Gemini-1.5, GPT-4V, GPT-4o, LLaMA-3.2-Vision and LLaVA-v1.6. Our code and data are publicly available at https://github.com/uscnlp-lime/VisBias.

VisFactor: Benchmarking Fundamental Visual Cognition in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) have demonstrated remarkable advancements in multimodal understanding; however, their fundamental visual cognitive abilities remain largely underexplored. To bridge this gap, we introduce VisFactor, a novel benchmark derived from the Factor-Referenced Cognitive Test (FRCT), a well-established psychometric assessment of human cognition. VisFactor digitalizes vision-related FRCT subtests to systematically evaluate MLLMs across essential visual cognitive tasks including spatial reasoning, perceptual speed, and pattern recognition. We present a comprehensive evaluation of state-of-the-art MLLMs, such as GPT-4o, Gemini-Pro, and Qwen-VL, using VisFactor under diverse prompting strategies like Chain-of-Thought and Multi-Agent Debate. Our findings reveal a concerning deficiency in current MLLMs' fundamental visual cognition, with performance frequently approaching random guessing and showing only marginal improvements even with advanced prompting techniques. These results underscore the critical need for focused research to enhance the core visual reasoning capabilities of MLLMs. To foster further investigation in this area, we release our VisFactor benchmark at https://github.com/CUHK-ARISE/VisFactor.

Libra-Leaderboard: Towards Responsible AI through a Balanced Leaderboard of Safety and Capability

To address this gap, we introduce Libra-Leaderboard, a comprehensive framework designed to rank LLMs through a balanced evaluation of performance and safety. Combining a dynamic leaderboard with an interactive LLM arena, Libra-Leaderboard encourages the joint optimization of capability and safety. Unlike traditional approaches that average performance and safety metrics, Libra-Leaderboard uses a distance-to-optimal-score method to calculate the overall rankings. This approach incentivizes models to achieve a balance rather than excelling in one dimension at the expense of some other ones. In the first release, Libra-Leaderboard evaluates 26 mainstream LLMs from 14 leading organizations, identifying critical safety challenges even in state-of-the-art models.

Difficult Task Yes but Simple Task No: Unveiling the Laziness in Multimodal LLMs

Multimodal Large Language Models (MLLMs) demonstrate a strong understanding of the real world and can even handle complex tasks. However, they still fail on some straightforward visual question-answering (VQA) problems. This paper dives deeper into this issue, revealing that models tend to err when answering easy questions (e.g. Yes/No questions) about an image, even though they can correctly describe it. We refer to this model behavior discrepancy between difficult and simple questions as model laziness. To systematically investigate model laziness, we manually construct LazyBench, a benchmark that includes Yes/No, multiple choice, short answer questions, and image description tasks that are related to the same subjects in the images. Based on LazyBench, we observe that laziness widely exists in current advanced MLLMs (e.g. GPT-4o, Gemini-1.5-pro, Claude 3 and LLaVA-v1.5-13B), and it is more pronounced on stronger models. We also analyze the VQA v2 (LLaVA-v1.5-13B) benchmark and find that about half of its failure cases are caused by model laziness, which further highlights the importance of ensuring that the model fully utilizes its capability. To this end, we conduct preliminary exploration on how to mitigate laziness and find that chain of thought (CoT) can effectively address this issue.

Insight Over Sight? Exploring the Vision-Knowledge Conflicts in Multimodal LLMs

This paper explores the problem of commonsense-level vision-knowledge conflict in Multimodal Large Language Models (MLLMs), where visual information contradicts model's internal commonsense knowledge (see Figure 1). To study this issue, we introduce an automated pipeline, augmented with human-in-the-loop quality control, to establish a benchmark aimed at simulating and assessing the conflicts in MLLMs. Utilizing this pipeline, we have crafted a diagnostic benchmark comprising 374 original images and 1,122 high-quality question-answer (QA) pairs. This benchmark covers two types of conflict target and three question difficulty levels, providing a thorough assessment tool. Through this benchmark, we evaluate the conflict-resolution capabilities of nine representative MLLMs across various model families and find a noticeable over-reliance on textual queries. Drawing on these findings, we propose a novel prompting strategy, "Focus-on-Vision" (FoV), which markedly enhances MLLMs' ability to favor visual data over conflicting textual knowledge. Our detailed analysis and the newly proposed strategy significantly advance the understanding and mitigating of vision-knowledge conflicts in MLLMs. The data and code are made publicly available.

Chain-of-Jailbreak Attack for Image Generation Models via Editing Step by Step

Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows, making them the focus in recent years. Despite their remarkable capability to generate diverse and vivid images, considerable efforts are being made to prevent the generation of harmful content, such as abusive, violent, or pornographic material. To assess the safety of existing models, we introduce a novel jailbreaking method called Chain-of-Jailbreak (CoJ) attack, which compromises image generation models through a step-by-step editing process. Specifically, for malicious queries that cannot bypass the safeguards with a single prompt, we intentionally decompose the query into multiple sub-queries. The image generation models are then prompted to generate and iteratively edit images based on these sub-queries. To evaluate the effectiveness of our CoJ attack method, we constructed a comprehensive dataset, CoJ-Bench, encompassing nine safety scenarios, three types of editing operations, and three editing elements. Experiments on four widely-used image generation services provided by GPT-4V, GPT-4o, Gemini 1.5 and Gemini 1.5 Pro, demonstrate that our CoJ attack method can successfully bypass the safeguards of models for over 60% cases, which significantly outperforms other jailbreaking methods (i.e., 14%). Further, to enhance these models' safety against our CoJ attack method, we also propose an effective prompting-based method, Think Twice Prompting, that can successfully defend over 95% of CoJ attack. We release our dataset and code to facilitate the AI safety research.

Learning to Ask: When LLMs Meet Unclear Instruction

Equipped with the capability to call functions, modern large language models (LLMs) can leverage external tools for addressing a range of tasks unattainable through language skills alone. However, the effective execution of these tools relies heavily not just on the advanced capabilities of LLMs but also on precise user instructions, which often cannot be ensured in the real world. To evaluate the performance of LLMs tool-use under imperfect instructions, we meticulously examine the real-world instructions queried from users, analyze the error patterns, and build a challenging tool-use benchmark called Noisy ToolBench (NoisyToolBench). We find that due to the next-token prediction training objective, LLMs tend to arbitrarily generate the missed argument, which may lead to hallucinations and risks. To address this issue, we propose a novel framework, Ask-when-Needed (AwN), which prompts LLMs to ask questions to users whenever they encounter obstacles due to unclear instructions. Moreover, to reduce the manual labor involved in user-LLM interaction and assess LLMs performance in tool utilization from both accuracy and efficiency perspectives, we design an automated evaluation tool named ToolEvaluator. Our experiments demonstrate that the AwN significantly outperforms existing frameworks for tool learning in the NoisyToolBench. We will release all related code and datasets to support future research.

On the Resilience of Multi-Agent Systems with Malicious Agents

Multi-agent systems, powered by large language models, have shown great abilities across various tasks due to the collaboration of expert agents, each focusing on a specific domain. However, when agents are deployed separately, there is a risk that malicious users may introduce malicious agents who generate incorrect or irrelevant results that are too stealthy to be identified by other non-specialized agents. Therefore, this paper investigates two essential questions: (1) What is the resilience of various multi-agent system structures (e.g., A$\rightarrow$B$\rightarrow$C, A$\leftrightarrow$B$\leftrightarrow$C) under malicious agents, on different downstream tasks? (2) How can we increase system resilience to defend against malicious agents? To simulate malicious agents, we devise two methods, AutoTransform and AutoInject, to transform any agent into a malicious one while preserving its functional integrity. We run comprehensive experiments on four downstream multi-agent systems tasks, namely code generation, math problems, translation, and text evaluation. Results suggest that the "hierarchical" multi-agent structure, i.e., A$\rightarrow$(B$\leftrightarrow$C), exhibits superior resilience with the lowest performance drop of $23.6\%$, compared to $46.4\%$ and $49.8\%$ of other two structures. Additionally, we show the promise of improving multi-agent system resilience by demonstrating that two defense methods, introducing an additional agent to review and correct messages or mechanisms for each agent to challenge others' outputs, can enhance system resilience. Our code and data are available at https://github.com/CUHK-ARISE/MAS-Resilience.