Reproduce “Using Experience Twice: Dynamic Process Rubrics for Privileged Token-Level Credit Assignment” 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
Shumao Wu, Kongcheng Zhang, Jianqiang Huang, Baisheng Lai
ORSD uses existing agent attempts to assign local training credit to tokens under sparse outcome feedback. It generates rubrics from trajectories for the same task, selects one whose criteria rankings align with rewards, then uses rubric-conditioned token likelihood changes to adjust GRPO advantages. On WebShop, Qwen3-4B reaches 77.05% success versus 66.15% for GRPO, without extra environment rollouts; this does not prove token correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.