EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
The paper’s core claim is that search queries should evolve alongside candidate solutions, not sit outside the optimization loop.
EvoDuet adds a retrieval gate that lets the model decide whether it needs new documents, stored documents, or no retrieval for the next step. Its inner loop refines queries and ranks documents by expected solution value, while the outer loop tests candidate solutions and feeds the results back into later searches. In 21 tasks, it improved normalized discovery gain for GPT-5.6-Luna and Gemini-3.8-Flash, but not Qwen3.5-9B. The authors report best-run gains over prior best scores on eight tasks and ties on three more. ArXiv · AI/CL/LG's note
EvoDuet adds a retrieval gate that lets the model decide whether it needs new documents, stored documents, or no retrieval for the next step. Its inner loop refines queries and ranks documents by expected solution value, while the outer loop tests candidate solutions and feeds the results back into later searches. In 21 tasks, it improved normalized discovery gain for GPT-5.6-Luna and Gemini-3.8-Flash, but not Qwen3.5-9B. The authors report best-run gains over prior best scores on eight tasks and ties on three more. ArXiv · AI/CL/LG's note
score 5