Reproduce “Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL” 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
Chuyi Tan, Peiwen Yuan, Xinglin Wang, Yiwei Li, Shaoxiong Feng, Yueqi Zhang, Jiayi Shi, Ji Zhang and 3 more
RLER trains reasoning policies on unlabeled queries by combining reward judgments from multiple policies, calibrated ensemble confidence, and selective rollout updates, then merging the policies for deployment. On DAPO-MATH-17K with Qwen2.5-Math-7B, it reaches 37.5 Avg@8 across six math benchmarks, above Co-rewarding’s 35.3 but below label-based RLVR’s 38.9 under these tested conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.