Youliang Yuan

Youliang Yuan, Wenxiang Jiao, Wenxuan Wang et al. · pku, tencent-ai

Safety lies at the core of the development of Large Language Models (LLMs). There is ample work on aligning LLMs with human ethics and preferences, including data filtering in pretraining, supervised fine-tuning, reinforcement learning from human feedback, and red teaming, etc. In this study, we discover that chat in cipher can bypass the safety alignment techniques of LLMs, which are mainly conducted in natural languages. We propose a novel framework CipherChat to systematically examine the generalizability of safety alignment to non-natural languages -- ciphers. CipherChat enables humans to chat with LLMs through cipher prompts topped with system role descriptions and few-shot enciphered demonstrations. We use CipherChat to assess state-of-the-art LLMs, including ChatGPT and GPT-4 for different representative human ciphers across 11 safety domains in both English and Chinese. Experimental results show that certain ciphers succeed almost 100% of the time to bypass the safety alignment of GPT-4 in several safety domains, demonstrating the necessity of developing safety alignment for non-natural languages. Notably, we identify that LLMs seem to have a ''secret cipher'', and propose a novel SelfCipher that uses only role play and several demonstrations in natural language to evoke this capability. SelfCipher surprisingly outperforms existing human ciphers in almost all cases. Our code and data will be released at https://github.com/RobustNLP/CipherChat.

Jen-tse Huang, Jiaxu Zhou, Tailin Jin et al. · allen-ai, cmu

Large language model-based multi-agent systems have shown great abilities across various tasks due to the collaboration of expert agents, each focusing on a specific domain. However, the impact of clumsy or even malicious agents--those who frequently make errors in their tasks--on the overall performance of the system remains underexplored. This paper investigates: (1) What is the resilience of various system structures (e.g., A$\rightarrow$B$\rightarrow$C, A$\leftrightarrow$B$\leftrightarrow$C) under faulty agents, on different downstream tasks? (2) How can we increase system resilience to defend against these agents? To simulate faulty agents, we propose two approaches--AutoTransform and AutoInject--which introduce mistakes into the agents' responses. Experiments on four downstream tasks using six systems show that the "hierarchical" structure, i.e., A$\rightarrow$(B$\leftrightarrow$C), exhibits superior resilience with the lowest performance drop of 5.5%, compared to 10.5% and 23.7% of other two structures. To further improve resilience, we introduce (1) Challenger, that introduces a mechanism for each agent to challenge others' outputs, and (2) Inspector, an additional agent to review and correct messages, recovering up to 96.4% errors made by faulty agents. Our code and data are available at https://github.com/CUHK-ARISE/MAS-Resilience.

Wenxuan Wang, Zhaopeng Tu, Chang Chen et al. · pku, tencent-ai

Safety lies at the core of developing and deploying large language models (LLMs). However, previous safety benchmarks only concern the safety in one language, e.g. the majority language in the pretraining data such as English. In this work, we build the first multilingual safety benchmark for LLMs, XSafety, in response to the global deployment of LLMs in practice. XSafety covers 14 kinds of commonly used safety issues across 10 languages that span several language families. We utilize XSafety to empirically study the multilingual safety for 4 widely-used LLMs, including both close-API and open-source models. Experimental results show that all LLMs produce significantly more unsafe responses for non-English queries than English ones, indicating the necessity of developing safety alignment for non-English languages. In addition, we propose several simple and effective prompting methods to improve the multilingual safety of ChatGPT by evoking safety knowledge and improving cross-lingual generalization of safety alignment. Our prompting method can significantly reduce the ratio of unsafe responses from 19.1% to 9.7% for non-English queries. We release our data at https://github.com/Jarviswang94/Multilingual_safety_benchmark.

Jen-tse Huang, Wenxuan Wang, Eric John Li et al. · pku, tencent-ai

Large Language Models (LLMs) have recently showcased their remarkable capacities, not only in natural language processing tasks but also across diverse domains such as clinical medicine, legal consultation, and education. LLMs become more than mere applications, evolving into assistants capable of addressing diverse user requests. This narrows the distinction between human beings and artificial intelligence agents, raising intriguing questions regarding the potential manifestation of personalities, temperaments, and emotions within LLMs. In this paper, we propose a framework, PsychoBench, for evaluating diverse psychological aspects of LLMs. Comprising thirteen scales commonly used in clinical psychology, PsychoBench further classifies these scales into four distinct categories: personality traits, interpersonal relationships, motivational tests, and emotional abilities. Our study examines five popular models, namely text-davinci-003, gpt-3.5-turbo, gpt-4, LLaMA-2-7b, and LLaMA-2-13b. Additionally, we employ a jailbreak approach to bypass the safety alignment protocols and test the intrinsic natures of LLMs. We have made PsychoBench openly accessible via https://github.com/CUHK-ARISE/PsychoBench.

Youliang Yuan, Wenxiang Jiao, Wenxuan Wang et al. · pku, tencent-ai

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses baseline methods in defending against attacks.

Wenxuan Wang, Juluan Shi, Zixuan Ling et al. · pku, tencent-ai

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.

Yusong Zhao, Yuejin Xie, Youliang Yuan et al.

Between the first visible sign of danger and the moment an accident occurs, there is often a window where intervention remains possible. Video-capable multimodal large language models (MLLMs) could serve as always-on safety monitors that issue warnings during this window. Yet current benchmarks do not test this ability: they rely on static inputs, ignore timing precision, and omit false-positive measurement on safe scenes. We present PaSBench-Video, a 740-video benchmark with 481 risk and 259 no-risk videos across four domains: driving, healthcare, daily life, and industrial production. Risk videos are annotated with frame-level risk onset and accident boundaries. A model must observe the video causally and produce a warning that is both temporally calibrated and content-correct. Testing 13 MLLMs, we find that no model exceeds 20.0% on our strictest metric, and recall is tightly coupled with false-positive rate, with Pearson correlation 0.64: higher detection comes only at the cost of triggering warnings on the majority of safe clips. Performance splits sharply by domain: models achieve moderate recall at low false-positive rates in daily life, where risks are inherently anomalous, yet fire indiscriminately in driving, where routine and hazardous scenes look alike. These results indicate that current models rely on scene-level activity cues rather than reasoning about emerging harm.

Sihang, Zhao, Kangrui Yu et al.

Large Language Models (LLMs) have been widely explored in educational scenarios. We identify a critical vulnerability in current educational LLMs, pedagogical jailbreaks, where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions. To enable systematic study, we unify and formalize safe, helpful, and pedagogical behaviors with a knowledge-mastery graph and introduce SHAPE, a benchmark of 9,087 student-question pairs for evaluating tutoring behavior under adversarial pressure. We propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries, identifies mastery gaps, and routes generation between instructing and problem-solving via explicit gating. Experiments across multiple LLMs show that our method yields significantly improved safety under two pedagogical jailbreak settings, while maintaining near-ceiling helpfulness under the same evaluation protocol. Our code and data are available at https://github.com/MAPS-research/SHaPE

Jen-tse Huang, Eric John Li, Man Ho Lam et al. · pku, tencent-ai

Decision-making is a complex process requiring diverse abilities, making it an excellent framework for evaluating Large Language Models (LLMs). Researchers have examined LLMs' decision-making through the lens of Game Theory. However, existing evaluation mainly focus on two-player scenarios where an LLM competes against another. Additionally, previous benchmarks suffer from test set leakage due to their static design. We introduce GAMA($γ$)-Bench, a new framework for evaluating LLMs' Gaming Ability in Multi-Agent environments. It includes eight classical game theory scenarios and a dynamic scoring scheme specially designed to quantitatively assess LLMs' performance. $γ$-Bench allows flexible game settings and adapts the scoring system to different game parameters, enabling comprehensive evaluation of robustness, generalizability, and strategies for improvement. Our results indicate that GPT-3.5 demonstrates strong robustness but limited generalizability, which can be enhanced using methods like Chain-of-Thought. We also evaluate 13 LLMs from 6 model families, including GPT-3.5, GPT-4, Gemini, LLaMA-3.1, Mixtral, and Qwen-2. Gemini-1.5-Pro outperforms others, scoring of $69.8$ out of $100$, followed by LLaMA-3.1-70B ($65.9$) and Mixtral-8x22B ($62.4$). Our code and experimental results are publicly available at https://github.com/CUHK-ARISE/GAMABench.

Yuxuan Wan, Wenxuan Wang, Yiliu Yang et al. · pku, tencent-ai

We introduce LogicAsker, a novel approach for evaluating and enhancing the logical reasoning capabilities of large language models (LLMs) such as ChatGPT and GPT-4. Despite LLMs' prowess in tasks like writing assistance, code generation, and machine translation, assessing their ability to reason has been challenging. Traditional evaluations often prioritize accuracy on downstream tasks over direct assessments of reasoning processes. LogicAsker addresses this gap by employing a set of atomic reasoning skills grounded in propositional and predicate logic to systematically examine and improve the reasoning prowess of LLMs. Our methodology reveals significant gaps in LLMs' learning of logical rules, with identified reasoning failures ranging from 29\% to 90\% across different models. Moreover, we leverage these findings to construct targeted demonstration examples and fine-tune data, notably enhancing logical reasoning in models like GPT-4o by up to 5\%. To our knowledge, this is the first effort to utilize test case outcomes to effectively refine LLMs' formal reasoning capabilities. We make our code, data, and results publicly available (https://github.com/yxwan123/LogicAsker) to facilitate further research and replication of our findings.

Jen-tse Huang, Jiantong Qin, Jianping Zhang et al. · pku, tencent-ai

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.

Jen-Tse Huang, Dasen Dai, Jen-Yuan Huang et al.

Despite significant progress on popular multimodal benchmarks, state-of-the-art Multimodal Large Language Models (MLLMs) continue to struggle with basic visual reasoning tasks that are trivially solved by humans, such as recognizing spatial relationships. To systematically investigate this gap, we introduce VisFactor, a benchmark that digitizes 20 vision-centric subtests from a well-established cognitive psychology assessment. These subtests span four core domains of human visual cognition: (1) Visualization and Spatial Processing, (2) Perceptual and Closure, (3) Memory, and (4) Reasoning. We evaluate 20 frontier MLLMs from GPT, Gemini, Claude, LLaMA, Qwen, and SEED families. The best-performing model achieves a score of only 25.19 out of 100, with consistent failures on tasks such as mental rotation, spatial relation inference, and figure-ground discrimination, regardless of model size or prompting strategy. These findings suggest that current MLLM performance gains on high-level benchmarks do not reflect human-like low-level visual cognition, challenging the assumption that large-scale pretraining naturally induces gestalt-like perceptual capabilities. The dataset and evaluation toolkit are publicly available at: https://github.com/CUHK-ARISE/VisFactor.

Youliang Yuan, Wenxiang Jiao, Yuejin Xie et al. · pku, tencent-ai

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.

Wenxuan Wang, Haonan Bai, Jen-tse Huang et al. · pku, tencent-ai

Image generation models can generate or edit images from a given text. Recent advancements in image generation technology, exemplified by DALL-E and Midjourney, have been groundbreaking. These advanced models, despite their impressive capabilities, are often trained on massive Internet datasets, making them susceptible to generating content that perpetuates social stereotypes and biases, which can lead to severe consequences. Prior research on assessing bias within image generation models suffers from several shortcomings, including limited accuracy, reliance on extensive human labor, and lack of comprehensive analysis. In this paper, we propose BiasPainter, a novel evaluation framework that can accurately, automatically and comprehensively trigger social bias in image generation models. BiasPainter uses a diverse range of seed images of individuals and prompts the image generation models to edit these images using gender, race, and age-neutral queries. These queries span 62 professions, 39 activities, 57 types of objects, and 70 personality traits. The framework then compares the edited images to the original seed images, focusing on the significant changes related to gender, race, and age. BiasPainter adopts a key insight that these characteristics should not be modified when subjected to neutral prompts. Built upon this design, BiasPainter can trigger the social bias and evaluate the fairness of image generation models. We use BiasPainter to evaluate six widely-used image generation models, such as stable diffusion and Midjourney. Experimental results show that BiasPainter can successfully trigger social bias in image generation models. According to our human evaluation, BiasPainter can achieve 90.8% accuracy on automatic bias detection, which is significantly higher than the results reported in previous work.

Wenxuan Wang, Xiaoyuan Liu, Kuiyi Gao et al. · pku, tencent-ai

Multimodal Large Language Models (MLLMs) have expanded the capabilities of traditional language models by enabling interaction through both text and images. However, ensuring the safety of these models remains a significant challenge, particularly in accurately identifying whether multimodal content is safe or unsafe-a capability we term safety awareness. In this paper, we introduce MMSafeAware, the first comprehensive multimodal safety awareness benchmark designed to evaluate MLLMs across 29 safety scenarios with 1500 carefully curated image-prompt pairs. MMSafeAware includes both unsafe and over-safety subsets to assess models abilities to correctly identify unsafe content and avoid over-sensitivity that can hinder helpfulness. Evaluating nine widely used MLLMs using MMSafeAware reveals that current models are not sufficiently safe and often overly sensitive; for example, GPT-4V misclassifies 36.1% of unsafe inputs as safe and 59.9% of benign inputs as unsafe. We further explore three methods to improve safety awareness-prompting-based approaches, visual contrastive decoding, and vision-centric reasoning fine-tuning-but find that none achieve satisfactory performance. Our findings highlight the profound challenges in developing MLLMs with robust safety awareness, underscoring the need for further research in this area. All the code and data will be publicly available to facilitate future research.

Haonan Li, Xudong Han, Zenan Zhai et al.

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.

Sihang Zhao, Youliang Yuan, Xiaoying Tang et al.

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.

Wenxuan Wang, Juluan Shi, Zhaopeng Tu et al. · pku, tencent-ai

Large Language Models (LLMs) like ChatGPT are foundational in various applications due to their extensive knowledge from pre-training and fine-tuning. Despite this, they are prone to generating factual and commonsense errors, raising concerns in critical areas like healthcare, journalism, and education to mislead users. Current methods for evaluating LLMs' veracity are limited by test data leakage or the need for extensive human labor, hindering efficient and accurate error detection. To tackle this problem, we introduce a novel, automatic testing framework, FactChecker, aimed at uncovering factual inaccuracies in LLMs. This framework involves three main steps: First, it constructs a factual knowledge graph by retrieving fact triplets from a large-scale knowledge database. Then, leveraging the knowledge graph, FactChecker employs a rule-based approach to generates three types of questions (Yes-No, Multiple-Choice, and WH questions) that involve single-hop and multi-hop relations, along with correct answers. Lastly, it assesses the LLMs' responses for accuracy using tailored matching strategies for each question type. Our extensive tests on six prominent LLMs, including text-davinci-002, text-davinci-003, ChatGPT~(gpt-3.5-turbo, gpt-4), Vicuna, and LLaMA-2, reveal that FactChecker can trigger factual errors in up to 45\% of questions in these models. Moreover, we demonstrate that FactChecker's test cases can improve LLMs' factual accuracy through in-context learning and fine-tuning (e.g., llama-2-13b-chat's accuracy increase from 35.3\% to 68.5\%). We are making all code, data, and results available for future research endeavors.

Changyue Li, Jiaying Li, Youliang Yuan et al.

Multimodal Large Language Models (MLLMs) are increasingly deployed in stateless systems, such as autonomous driving and robotics. This paper investigates a novel threat: Semantic-Aware Hijacking. We explore the feasibility of hijacking multiple stateless decisions simultaneously using a single universal perturbation. We introduce the Semantic-Aware Universal Perturbation (SAUP), which acts as a semantic router, "actively" perceiving input semantics and routing them to distinct, attacker-defined targets. To achieve this, we conduct theoretical and empirical analysis on the geometric properties in the latent space. Guided by these insights, we propose the Semantic-Oriented (SORT) optimization strategy and annotate a new dataset with fine-grained semantics to evaluate performance. Extensive experiments on three representative MLLMs demonstrate the fundamental feasibility of this attack, achieving a 66% attack success rate over five targets using a single frame against Qwen.

Youliang Yuan, Qiuyang Mang, Jingbang Chen et al. · pku, tencent-ai

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.