EvoDuet Co-Evolves Queries and Scientific Solutions

EvoDuet keeps weights fixed and co-evolves web queries with solutions, raising discovery scores on 21 tasks for two of three models.

EvoDuet is a bilevel method that co-evolves scientific-discovery solutions and web-search queries without updating model parameters. Both reports describe a retrieval gate that can fetch new documents, reuse stored ones, or skip retrieval, and an inner loop that ranks documents by predicted solution score. Across 21 tasks with one candidate per iteration, it lifts OpenEvolve's normalized discovery gain from 74.1% to 78.0% with GPT-5.6-Luna and from 61.3% to 82.3% with Gemini-3.8-Flash; both say Qwen3.5-9B does not improve. Best runs beat previously reported scores on eight tasks. The sources differ on extras: one names Swap Reduction on Q20 and Rosetta among those eight and says three more tasks are matched, while the other says the inner loop refines queries, an outer loop scores candidates in parallel, and the method also helps other evolutionary scaffolds. They do not disagree on the headline percentages.

• EvoDuet co-evolves candidate solutions and web-search queries while keeping model weights fixed.
• A retrieval gate can fetch new documents, reuse stored ones, or proceed without retrieval; an inner loop ranks documents by predicted solution score.
• On 21 tasks with one candidate per iteration, OpenEvolve normalized discovery gain rises from 74.1% to 78.0% with GPT-5.6-Luna and from 61.3% to 82.3% with Gemini-3.8-Flash.
• Both reports say Qwen3.5-9B does not benefit.
• Best runs exceed previously reported best scores on eight of 21 tasks; one report names Swap Reduction on Q20 and Rosetta and says three further tasks are matched.
• One report says the inner loop refines queries and an outer loop generates and scores candidates in parallel, and that the method also helps other evolutionary scaffolds.