Reproduce “BLPO: Beyond-Local Policy Optimization for Long-Horizon LLM Agents” 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

Xuanren Chen, Chongyang Tao, Wenpeng Hu, Yong Wang, Bo Deng, Xinlei Zheng, Shuai Ma

BLPO trains long-horizon agents by sharing comparisons across tasks at matching progress stages and action functions. It adjusts for context difficulty and weak evidence, while retaining local and trajectory-level credit without a learned critic or extra rollouts. In evaluated ALFWorld and WebShop settings, it improves task success over baselines; same-task-only memory scores lower than full BLPO on WebShop.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.