HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses

HypoEvolve. Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses. By Jieyuan Liu and twelve others, from the University of California San Diego, Texas A and M University, and two other institutions. Published on the fourteenth of September. No code repository is linked to it.

Scientific agents help discover hypotheses by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems pair such agents with evolutionary search, through critique, comparison and revision. But one question is still open: how do different forms of collaboration between agents affect the quality of the hypotheses? Answering it means separating two things that are usually tangled together: how good the agents are at science, and how they collaborate. A framework built for that has to preserve the agents' scientific roles, and support explicit rules for combining, revising, and retaining hypotheses. Without that separation, a better result could be credited to collaboration when it really came from stronger individual agents, or the other way round.

HypoEvolve makes collaboration explicit, as a series of updates to a population of hypotheses. A generational genetic algorithm coordinates specialized large language model agents. Some integrate mechanistic arguments, some reconsider assumptions, and some assess evidence and testability. Each generation spells out how the scientific judgements and the new proposals reshape the population. Because those rules are explicit, the effect of collaboration on hypothesis quality becomes something that can be tested directly. Working in generations suits the question. Each generation is a round of combining, revising, and retaining hypotheses, the very rules the framework was required to support. Here each of those steps is carried out by scientific agents with defined roles, so each step is also a place where the form of collaboration can be varied and its effect observed.

The evaluation is designed around hypotheses that are scientifically meaningful: explanations of how a proposed intervention could work. Drug repurposing is the testbed, because it ties those explanations to target-level biological claims that can be checked against external evidence. The authors adapt DepMap and Open Targets into two complementary external measures, grounded in experimental, genetic and clinical evidence. Using outside evidence matters for the design. The hypotheses are not judged only by other language models, but against measures built from what experiments, genetics, and clinical data already record.

Across thirty-four cancer types, HypoEvolve achieves the highest scores on both measures against six baselines. Its DepMap selectivity reaches zero point one seven one, against zero point one one five for the strongest baseline. Its gains over single-pass generation also generalize to cancer types that were held out. Held-out generalization is the check that the gains are not tied to the particular cancer types the system was developed on.

The larger claim is about teams. The authors see HypoEvolve as a step toward autonomous science in which AI research teams reach a capacity for discovery beyond that of any individual model. What makes that claim testable here is the genetic algorithm: it turns collaboration from an informal exchange into a set of rules whose effect on the population can be measured.