Reproduce “PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models” 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
Yuxuan Chen, Xiaojie Li, Hao Chen, Juncheng Ma, Kewei Zhang, Zhongjing Du, Guoqiang Wei, Fangfang Wang and 6 more
PhysLoop uses executable simulations to construct and check physical events, then refines videos for training examples or localized repair. It verifies simulated motion against task conditions and checks refined outputs, with feedback guiding revisions. On a 50-task repair evaluation, the full system achieves 86.00% success versus 19.33% for standalone Seedance 2.5. This is task-specific performance, not proof of real-world physical accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.