EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery

The work introduces EvoDuet, a bi-level optimization method that co-evolves solutions and search queries using fixed model parameters to prevent evolutionary search stalls. A retrieval gate allows the LLM to assess knowledge gaps, while an inner loop refines queries and ranks documents based on predicted solution scores, and an outer loop generates parallel candidates. Across 21 optimization tasks, EvoDuet 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, while surpassing prior best scores on eight tasks.

EvoDuet raises OpenEvolve's normalized discovery gain from 74.1% to 78.0% with GPT-5.6-Luna.

It increases the gain from 61.3% to 82.3% with Gemini-3.8-Flash, whereas Qwen3.5-9B does not benefit.

It surpasses previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta.