Researchers have introduced HypoEvolve, a framework that uses a generational genetic algorithm to coordinate specialized large language model (LLM) agents in the discovery of scientific hypotheses. The work, published as a research paper, addresses an open question in AI-driven science: how different forms of agent collaboration affect the quality of generated hypotheses.
According to the paper, answering this question requires separating the effects of agents' scientific capabilities from those of their collaboration. HypoEvolve therefore makes collaboration explicit through successive updates to a hypothesis population. Each generation specifies how scientific judgments and new proposals reshape that population, making the effects of collaboration on hypothesis quality directly testable.
The framework coordinates agents that integrate mechanistic arguments, reconsider assumptions, and assess evidence and testability. The authors designed their evaluation around scientifically meaningful hypotheses that explain how a proposed intervention could work, using drug repurposing to link these explanations to target-level biological claims assessed against external evidence.
Specifically, the team adapted DepMap and Open Targets into complementary external measures grounded in experimental, genetic, and clinical evidence. Across 34 cancer types, HypoEvolve achieved the highest scores against six baselines on both measures. DepMap selectivity reached 0.171, compared with 0.115 for the strongest baseline. The authors report that gains over single-pass generation also generalize to held-out cancer types.
The paper frames HypoEvolve as advancing a vision of autonomous science in which AI research teams achieve a capacity for discovery beyond that of individual models. The authors emphasize that the key idea is to make collaboration itself something that can be systematically designed and improved.