Reproduce “Verifier-Induced Support Reshaping in On-Policy Optimization” 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
Shaohang Wei, Zikun Su, Feifan Song, Wen Luo, Wei Li, Guangyue Peng, Houfeng Wang
The paper examines how reinforcement learning with verifiable rewards can improve one training objective while reducing successful responses available to a later objective. Across Qwen models and instruction-following benchmarks, Math-RLVR raises one-rollout success, pass@1, but lowers 32-sample prompt coverage, best@32. On Qwen3-8B-Base IFEval, changes are +0.065 and −0.098. Opening interventions support a causal role, not erased ability or universal effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.