HypoEvolve wraps a genetic algorithm around multi-agent LLMs and beats six baselines on drug-repurposing hypothesis generation

arXiv·medium signalXBlueskyLinkedInCopy link

A September 14 arXiv paper from 13 authors runs a generational genetic algorithm over specialized LLM agents that separately handle mechanistic argument, assumption reconsideration, and evidence and testability assessment. On drug repurposing evaluated against DepMap and Open Targets across 34 cancer types, it reached 0.171 DepMap selectivity versus 0.115 for the strongest baseline, outperforming six baselines on both measures. Gains over single-pass generation also held on held-out cancer types, which is the part that matters for builders: the evolutionary loop, not the agent specialization alone, carried the improvement.