Bowen Song

Yiming Zhang, Yingfan Ma, Yanmei Gu et al.

Large Language Models (LLMs) have shown impressive performance in domains such as mathematics and programming, yet their capabilities in physics remain underexplored and poorly understood. Physics poses unique challenges that demand not only precise computation but also deep conceptual understanding and physical modeling skills. Existing benchmarks often fall short due to limited difficulty, multiple-choice formats, and static evaluation settings that fail to capture physical modeling ability. In this paper, we introduce ABench-Physics, a novel benchmark designed to rigorously evaluate LLMs' physical reasoning and generalization capabilities. ABench-Physics consists of two components: Phy_A, a static set of 400 graduate- or Olympiad-level problems; and Phy_B, a dynamic subset of 100 problems equipped with an automatic variation engine to test model robustness across changing conditions. All questions require precise numerical answers, with strict formatting and tolerance constraints. Our evaluation of several state-of-the-art LLMs reveals substantial performance gaps, highlighting persistent limitations in physical reasoning, especially in generalization to dynamic variants. ABench-Physics provides a challenging and diagnostic framework for advancing scientific reasoning in LLMs.

Bonan Zhang, Zhongqi Chen, Bowen Song et al.

Reinforcement learning (RL) has become a standard paradigm for refining large language models (LLMs) beyond pre-training and instruction tuning. A prominent line of work is RL with verifiable rewards (RLVR), which leverages automatically verifiable outcomes (e.g., correctness or executability) to generate reward signals. While efficient, this framework faces two key limitations: First, its binary feedback is too sparse to capture the quality of the reasoning process. Second, its coarse-grained rewards potentially lead to vanishing gradients. Inspired by observations from human learning, we introduce a RL technique that integrates verifiable outcomes with the model's own confidence estimates. This joint design enriches the reward signal, providing finer-grained feedback and implicitly supervising the reasoning process. Experimental results demonstrate that our proposed method enhances RL performance across multiple datasets and reduces token consumption during inference, while incurring negligible additional training cost. Moreover, it can be used as a plug-in module to enhance other state-of-the-art RL methods.