HypoEvolve and Iterative Hypothesis Refinement

HypoEvolve moves away from standard one-off model prompts toward a multi-agent framework that iteratively refines biological hypotheses. In a typical single-pass setup, an LLM generates a target and stops. HypoEvolve instead maintains a population of potential interventions, using specialized agents to propose mechanisms, critique their logic, and verify findings against databases like DepMap and Open Targets. The system replaces static completion with a cycle of successive updates where each hypothesis is treated as an evolving artifact.

Iterative Refinement Through Feedback

Rather than applying formal genetic operators, the system uses a cycle of successive updates. A generator agent drafts a mechanistic argument, which a critic agent then audits for logical gaps or unsupported claims. When the critic identifies an issue, the system passes the original text and the feedback back to the generator. The model is prompted to refine only the specific segment identified as problematic while keeping the surrounding context intact. This directed editing prevents the model from hallucinations common in single-pass generation by forcing it to anchor its output to the critic's report.

Instead of merging two hypotheses into a third, the framework maintains the population by selecting the most promising candidates from the current set to serve as the basis for the next iteration. This ensures the reasoning patterns that proved successful in previous cycles are preserved as foundations for further exploration.

Assessment and Grounding

To manage this population, the system uses an assessment mechanism that checks hypotheses against known biological data. When the model retrieves records from sources like DepMap, it evaluates whether the data supports the causal link between a gene target and a specific cell line outcome. Hypotheses that align with retrieved evidence are prioritized for further development. If the retrieval process returns no supporting records, the hypothesis is deprioritized, which effectively prunes the population of unsupported ideas.

State Management and Novelty

Implementing this approach requires maintaining a persistent state where hypothesis strings are tracked alongside their retrieved evidence. This allows the agents to operate on a consistent history of updates rather than treating each prompt as an isolated task. The main constraint of this framework is its reliance on existing records to validate success. Because the scoring is tied directly to the presence of supporting evidence in the selected databases, the system does not distinguish between a truly novel, biologically valid hypothesis and a false one. In the current iteration, candidates without database matches are assigned low scores and are likely to be eliminated in favor of entries with established evidence, meaning the system is optimized for retrieval-backed plausibility rather than experimental discovery of entirely new phenomena.