Reproduce “How Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies” 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
Bowei He, Yankai Chen, Xiaokun Zhang, Changjiang Han, Ye Yuan, Chong Li, Xue Liu
The authors measure tasks, environments, and trajectories separately to guide compute allocation for agentic reinforcement-learning training. They fit scaling laws and synthesis costs, then test the resulting recipe on Qwen3-14B at 32× compute on a math-and-code testbed. It achieves 51.3% held-out task success, versus 47.3% and 42.9% for two baselines, in a single model-size and budget validation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.