Reproduce “REINFORCE Zero: Kernel Alignment Unlocks Stable Asynchronous RL with Importance Sampling” 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
Jian Hu, Hao Zhang, Yifan Zhang, Binfeng Xu, Ao Shen, Shaokun Zhang, Yunheng Zou, Yan Bai and 2 more
REINFORCE Zero uses asynchronous reinforcement learning to overlap generation and optimization. It aligns rollout and learner execution at identical weights and histories, then applies unclipped token-level importance sampling to correct weight lag without rejecting samples. On DeepSeek-R1-Distill-Qwen-1.5B, it improves through 20,000 updates at nominal lags 2 and 8, reaching five-benchmark macro pass@1 scores of 51.10% and 50.93% in this evaluation setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.