Reproduce “Libra: Efficient Resource Management for Agentic RL Post-Training” 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
Kaiwen Chen, Xin Tan, Jingzong Li, Zhi Zhou, Cen Li, Jiang Liu, Meng Jie, Jiazhi Jiang and 1 more
Libra addresses agentic RL bottlenecks by routing tool-using trajectories among rollout workers and reallocating workers between rollout and training. It uses observed tool returns to estimate remaining length and periodically replans resource splits. In a Search-R1 experiment, Qwen3-14B reached 2,900 tokens/s versus DynaRL’s 2,000, and target reward in 16.2 versus 23.5 hours. Results are limited to tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.