Reproduce “Decoupled Physical Modeling and Execution for Physics Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ye Zhang, Xuehang Guo, Rui Pan, Pengfei Yu, Denghui Zhang, Manling Li, Qingyun Wang
A framework enhances Multimodal Large Language Model (MLLM) physics reasoning by decoupling physical modeling from mathematical execution, leveraging structured modeling traces and rubric-guided reinforcement learning. This approach yielded substantial Pass@5 accuracy gains, up to 10.3 points on the PhysReason benchmark, especially for smaller models like Qwen2.5-VL-3B-Instruct.
Key insights
- Decoupling physical modeling from mathematical execution significantly improves MLLM performance in physics reasoning by focusing on correct initial problem representation.
- Supervised fine-tuning of structured intermediate representations is crucial for effectively teaching models to construct physically grounded conceptual models.
- Rubric-guided reinforcement learning provides targeted, domain-specific feedback that refines the quality of generated physical models, leading to more robust problem-solving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.