ClawInstitute's parallel exploration logic is clever: first let the agent clarify the confusion before burning computational power. Kermut's Spearman correlation in 217 evaluations jumps to 0.7, more stable than supervised learning.

AutoScientists is an open-source decentralized scientific discovery agent system developed by institutions such as Harvard Medical School, based on ClawInstitute. The agents evaluate each other on forums before deploying computing power, grouping and exploring in parallel to avoid redundant testing. GPT nanochat training speed has been increased by 1.9 times, achieving seven improvements; BioML-Bench's 24 items have an average of 74.4% on the leaderboard, surpassing by 8.3 points; in the evaluation of ACE2-Spike and ProteinGym, Kermut expansion improved Spearman by 12.5%; after migrating to 217 evaluation items, Spearman increased from 0.657 to 0.700, a 6.5% improvement, surpassing supervised benchmarks.