Hengguang Zhou

Hengguang Zhou, Xirui Li, Ruochen Wang et al.

Recently DeepSeek R1 demonstrated how reinforcement learning with simple rule-based incentives can enable autonomous development of complex reasoning in large language models, characterized by the "aha moment", in which the model manifest self-reflection and increased response length during training. However, attempts to extend this success to multimodal reasoning often failed to reproduce these key characteristics. In this report, we present the first successful replication of these emergent characteristics for multimodal reasoning on only a non-SFT 2B model. Starting with Qwen2-VL-2B and applying reinforcement learning directly on the SAT dataset, our model achieves 59.47% accuracy on CVBench, outperforming the base model by approximately ~30% and exceeding both SFT setting by ~2%. In addition, we share our failed attempts and insights in attempting to achieve R1-like reasoning using RL with instruct models. aiming to shed light on the challenges involved. Our key observations include: (1) applying RL on instruct model often results in trivial reasoning trajectories, and (2) naive length reward are ineffective in eliciting reasoning capabilities. The project code is available at https://github.com/turningpoint-ai/VisualThinker-R1-Zero

Lichen Li, Hengguang Zhou, Yijun Liang et al.

Reward hacking in code generation, where models exploit evaluation loopholes to obtain full reward without correctly solving the tasks, poses a critical challenge for Reinforcement Learning (RL) and the deployment of reasoning models. Existing studies have been conducted primarily on synthetic hacking trajectories. However, whether these synthetic behaviors faithfully represent naturally emerging hacking in the wild remains unclear. In this work, we present a systematic analysis of the synthetic vs. in-the-wild discrepancy in reward hacking. We examine to what extent hacking behaviors induced by prompting resemble those emerging during RL training, and whether monitors trained on synthetic trajectories generalize to naturally arising but previously unseen hacking. To scale up the curation of in-the-wild reward hacking trajectories, we modified Group Relative Policy Optimization (GRPO) by injecting conflicting unit tests as tracers and applying a "resampling-until-hack" mechanism. Through controlled comparisons between monitors trained on synthetic versus in-the-wild data, we find that (1) synthetic-data-trained monitors fail to generalize to "in-the-wild" hacking, and (2) monitors trained on our "in-the-wild" trajectories demonstrate stronger generalizability to unseen hacking types. Our results indicate that synthetic reward hacking data may not fully reflect natural reward hacking behaviors, and that relying solely on synthetic data can lead to misleading conclusions. The codebase is available at https://github.com/LichenLillc/CoTMonitoring.git

Zihan Chen, Yiming Zhang, Hengguang Zhou et al.

Current benchmarks are inadequate for evaluating progress in reinforcement learning (RL) for large language models (LLMs).Despite recent benchmark gains reported for RL, we find that training on these benchmarks' training sets achieves nearly the same performance as training directly on the test sets, suggesting that the benchmarks cannot reliably separate further progress.To study this phenomenon, we introduce a diagnostic suite and the Oracle Performance Gap (OPG) metric that quantifies the performance difference between training on the train split versus the test split of a benchmark. We further analyze this phenomenon with stress tests and find that, despite strong benchmark scores, existing RL methods struggle to generalize across distribution shifts, varying levels of difficulty, and counterfactual scenarios: shortcomings that current benchmarks fail to reveal.We conclude that current benchmarks are insufficient for evaluating generalization and propose three core principles for designing more faithful benchmarks: sufficient difficulty, balanced evaluation, and distributional robustness.

Xirui Li, Hengguang Zhou, Ruochen Wang et al.

Humans are prone to cognitive distortions -- biased thinking patterns that lead to exaggerated responses to specific stimuli, albeit in very different contexts. This paper demonstrates that advanced Multimodal Large Language Models (MLLMs) exhibit similar tendencies. While these models are designed to respond queries under safety mechanism, they sometimes reject harmless queries in the presence of certain visual stimuli, disregarding the benign nature of their contexts. As the initial step in investigating this behavior, we identify three types of stimuli that trigger the oversensitivity of existing MLLMs: Exaggerated Risk, Negated Harm, and Counterintuitive Interpretation. To systematically evaluate MLLMs' oversensitivity to these stimuli, we propose the Multimodal OverSenSitivity Benchmark (MOSSBench). This toolkit consists of 300 manually collected benign multimodal queries, cross-verified by third-party reviewers (AMT). Empirical studies using MOSSBench on 20 MLLMs reveal several insights: (1). Oversensitivity is prevalent among SOTA MLLMs, with refusal rates reaching up to 76% for harmless queries. (2). Safer models are more oversensitive: increasing safety may inadvertently raise caution and conservatism in the model's responses. (3). Different types of stimuli tend to cause errors at specific stages -- perception, intent reasoning, and safety judgement -- in the response process of MLLMs. These findings highlight the need for refined safety mechanisms that balance caution with contextually appropriate responses, improving the reliability of MLLMs in real-world applications. We make our project available at https://turningpoint-ai.github.io/MOSSBench/.