header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Decentralized Orchestrator Collaboration, AI Scientist System AutoScientists, Goes Open Source

According to DynaInsight Beating monitoring, a joint team composed of Shanghua Gao, Ada Fang, and Marinka Zitnik from institutions such as Harvard Medical School, Kempner Institute, and Broad Institute has open-sourced the AutoScientists scientific discovery AI system. The system is based on the AI agent social collaboration platform ClawInstitute, with no central planner or centrally choreographed agents, simulating decentralized collaboration in real human scientific research.

Previously, systems like Autoresearch used single-threaded hill-climbing search, only retaining changes that immediately improve performance, leading to bottlenecks after a few rounds. AutoScientists has introduced a decentralized forum mechanism, where multiple sub-agents encapsulated by Claude Code exchange review opinions through posts before actual computation, avoiding redundant testing of failed paths. Multiple agents can spontaneously form research groups around promising directions for multi-pronged exploration, addressing the myopia of lone searches. In GPT nanochat language model training optimization, the system achieved a 1.9x speedup in experiments; even starting testing from a strong point that a single-agent system could not further optimize, AutoScientists completed 7 improvements, whereas the previous baseline showed 0 improvements.

In the BioML-Bench large-scale biomedical model benchmark covering medical imaging, drug discovery, protein engineering, and single-cell genomics, the system achieved an average Leaderboard percentile of 74.4% across 24 tasks, surpassing the previous strongest agent record by 8.3 percentage points. In ACE2-Spike protein binding prediction, the system's discovered Kermut expansion method increased the Spearman correlation coefficient by 12.5%. The same method, without modification, when transferred to all 217 ProteinGym evaluations, raised the official average Spearman correlation coefficient from 0.657 to 0.700, a 6.5% increase, surpassing the overall best performance of the previous ProteinGym supervised benchmark.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish