Recursive self-improving agent MolRSI lifts wet-lab hit rate 82%, beats ClaudeScience 8-0
机器之心 · wechat · 2026-11-14
Researchers unveiled MolRSI (arXiv:2610.39714), a recursive self-improving AI agent for molecular discovery built on dual loops: an inner loop optimizes the current workflow via experience-guided Monte Carlo tree search, while an outer loop carries workflow, edit, and execution memories into future tasks.
Antibacterial screening
- From a 12.1M compound library, MolRSI picked 107 candidates for wet-lab testing and hit 17 (15.9%); the human-expert non-agentic workflow hit 21/241 (8.7%) — an 82% relative improvement under identical data and MSSA criteria.
- On an independent test set its activity predictor improved AUPRC from 0.090 to 0.300; representative molecules C3/C32 showed MRSA MICs of 1 and 8 μg/mL.
- Inherited experience matters: with 25 source tasks, its first evaluated workflow beat final workflows that Biomni, STELLA, Codex, ClaudeCode and others found after 20 search rounds.
De novo pMHC binder design
- On MART-1 and SARS-CoV-1 pMHC targets, phage display validation yielded 8 specific binding hits vs 0 for ClaudeScience at equal compute; removing candidate ranking produced zero hits out of 479 designs, confirming ranking's crucial role.
More from coding & agent
- kalomaze: 'subagents' are really invisible model calls, and one just rejected a completed task — kalomaze · 2026-11-14
- "You Too Will Use Claude Code": AI Coding Evangelism Gets Meme'd — max_paperclips · 2026-11-14
- Subagent misreads an abbreviation, stop hook blocks Claude Code after goal is met — kalomaze · 2026-11-14
- Dev shares 6-layer prompt pipeline with dual FAISS indexes to keep AI personas stable across models — OrionForgeEcosystem · 2026-11-14
- 'Claude unlocked Mac gaming': devs are porting games to macOS with AI and it works — sandeepssrin · 2026-11-14
- Running DeepSeek 4.1 Flash on a 32GB Mac via SSD streaming with open-source Local Code — alexcovo_eth · 2026-11-14