HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
The paper claims an evolutionary loop of specialized LLM agents produced stronger drug-repurposing hypotheses than six baselines across 34 cancer types.
HypoEvolve keeps a population of hypotheses and updates it generation by generation through agent critique, revision, retention, and new proposals. The agents are assigned scientific roles around mechanism, assumptions, evidence, and testability. The authors evaluate the system on intervention hypotheses tied to target-level biological claims, using adapted DepMap and Open Targets measures. Its DepMap selectivity score is reported at 0.171, compared with 0.115 for the strongest baseline, with gains also seen on held-out cancer types.
ArXiv · AI/CL/LG's note
HypoEvolve keeps a population of hypotheses and updates it generation by generation through agent critique, revision, retention, and new proposals. The agents are assigned scientific roles around mechanism, assumptions, evidence, and testability. The authors evaluate the system on intervention hypotheses tied to target-level biological claims, using adapted DepMap and Open Targets measures. Its DepMap selectivity score is reported at 0.171, compared with 0.115 for the strongest baseline, with gains also seen on held-out cancer types.
ArXiv · AI/CL/LG's note
score 4