Reproduce “When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing” 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
Lei Zhao, Zongheng Cao, Xinyu Hu
The authors compare trajectory-level outcome rewards, turn-level process rewards, and token-level self-distillation for learning movie-trailer editing. Turn-level rewards guide individual decisions, while outcome-only GRPO uses the final result. Decayed process rewards score 0.293 on Fine Cut, versus 0.201 for outcome-only GRPO, a reported 45.4% gain. The result is limited to this tested movie-trailer setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.