HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
HypoEvolve uses a genetic algorithm to coordinate specialized LLM agents and test how their collaboration changes hypothesis quality.
The system maintains a population of hypotheses, then updates it across generations through critique, revision, selection, and new proposals. Its evaluation focuses on drug-repurposing hypotheses for cancer, tying explanations to target-level biological claims. Across 34 cancer types, the paper reports stronger results than six baselines on DepMap and Open Targets measures, including DepMap selectivity of 0.171 versus 0.115 for the strongest baseline. HF Daily Papers' note
The system maintains a population of hypotheses, then updates it across generations through critique, revision, selection, and new proposals. Its evaluation focuses on drug-repurposing hypotheses for cancer, tying explanations to target-level biological claims. Across 34 cancer types, the paper reports stronger results than six baselines on DepMap and Open Targets measures, including DepMap selectivity of 0.171 versus 0.115 for the strongest baseline. HF Daily Papers' note
score 5