Reproduce “Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards” 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
Keru Chen, Sen Lin, Yingbin Liang, Nathaniel D. Bastian, Shaofeng Zou
CoSIR trains agents to share evidence, request missing facts, and use replies through rewards tied to information value. In a verifiable toy setting, it scores communication by whether information helps teammates identify the correct option. It has the highest aggregate score on seven unseen collaboration tasks for each of three backbones, with some sharing behaviors increasing; this does not establish transfer to every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.