MoLRSI: Recursive Self-Improving AI Agent Doubles Antibiotic Hit Rate to 15.9% in Wet Lab
LucaAmb · x · 2026-11-07
Wengong Jin's team released MoLRSI, a recursively self-improving AI agent for molecular discovery (arXiv 2610.39714).
The core idea: beyond automating expert-defined molecular design protocols, the agent systematically optimizes its workflows and model architectures, using persistent cross-task memory to recursively improve its search policy—workflow performance, refinement outcomes, and execution feedback all guide subsequent searches.
Wet-lab validated results:
- Antibiotic discovery: 15.9% experimental hit rate vs 8.7% for a prior expert-led workflow using the same training data and screening library, while balancing antibacterial activity, cytotoxicity, and structural novelty
- De novo pMHC binder design: 8 selective binder hits across 2 targets, vs 0 for Claude Science
Related event: Self-Improving AI Agent MoLRSI Doubles Antibiotic Hit Rate(2 posts)→
More from AGI Musings
- DeepMind's AGI economics program: millions of agents may transform economy with just 4-5% GDP growth — soumitrashukla9 · 2026-11-07
- Martin Casado sparks debate over lines of code as a metric for AI recursive self-improvement — wavefnx · 2026-11-07
- What happens if RSI-complete model weights leak: thousands of self-improvement loops at once — connoraxiotes · 2026-11-07
- "The World Is Not Ready for Post-Labor Economics": Aztec Meme Nails AI Anxiety — Chris_Armstrong · 2026-11-07
- "Agents Coordinating Flesh Tubes for Coffee": The Paternalism Problem of AI-Managed Teams — sarahdrinkwater · 2026-11-07
- ETH Zürich assigns unsolved open problems as AI makes textbook exercises obsolete — JeremyNguyenPhD · 2026-11-07