EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
EvoDuet co-evolves candidate solutions and web-search queries to improve LLM scientific discovery.
EvoDuet keeps model parameters fixed and co-evolves candidate solutions with web-search queries. A retrieval gate lets the model fetch new documents, reuse stored ones, or proceed without retrieval, while an inner loop ranks documents by predicted solution scores. Across 21 tasks with one candidate per iteration, it raises 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; Qwen3.5-9B does not benefit. The best runs exceed previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three more.
44