AutoScientists has open-sourced, delivering a 74.4% leaderboard percentile and 1.9x training acceleration, setting a new SOTA on Biomed-Bench. The Gao Shanghua team has finally figured out how to make AI Agents’ collaboration mechanism work—forum-style reviews are far more efficient than repeatedly stumbling into the same pitfalls.

Harvard Medical School and other institutions' Gao Shanghua and others developed and open-sourced AutoScientists, a decentralized collaboration platform for AI agents based on ClawInstitute, mimicking real scientific research with distributed teams. Unlike single-threaded hill-climbing methods, AutoScientists allows sub-intelligent agents to post and review each other on a decentralized forum before actual computing power is used, avoiding repeated testing of failed paths. The system achieved a 1.9x acceleration in training optimization on GPT NanoChat and attained an average leaderboard percentile of 74.4% across 24 tasks on the Biomed-Bench benchmark, surpassing previous best records.