Reproduce “Learning to Accumulate Knowledge with Mutual Information” 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

Yuyang Zhao, Lizi Liao, Leyang Shen, Xiaoyan Zhao, Yang Zhang, Fuli Feng, Xiangnan He

Knowledge Weaver trains a curator to create notes while the executor stays frozen. It rewards standalone success, context-specific tokens, and gains over retrieved guidance. With ten entries and a frozen Qwen3.5-4B executor, three-seed mean success rises from 37.1% to 54.0% on ALFWorld and 23.3% to 42.0% on WebShop. Results cover two text environments, not continual accumulation or joint executor training.

Summary by alphaXiv.

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