Researchers have introduced EvoDuet, a bi-level optimization method that co-evolves web search queries and candidate solutions while keeping the underlying language model's parameters fixed. The work targets a common failure mode in evolutionary search with large language models: progress stalls when a task requires external knowledge the model does not have.
Simply bolting a web search tool onto the process is not enough, according to the paper, because the same pages can keep resurfacing even as the candidate solutions evolve. EvoDuet addresses this with a retrieval gate at each iteration that lets the LLM assess its own knowledge gap and choose one of three paths: retrieve new documents, reuse stored ones, or proceed without retrieval.
The method splits the work into two loops. An inner loop refines queries and ranks documents by the solution scores they are predicted to produce, while an outer loop generates candidates in parallel from those documents and records the evaluated outcomes to inform later searches.
Across 21 optimization tasks with one candidate per iteration, EvoDuet raised 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. The smaller Qwen3.5-9B did not benefit. The authors report that their best runs surpass previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three others.
The approach also improved results with other evolutionary scaffolds, such as Top-K and EvoX, on the Sums/Diffs and Denoising tasks, which the authors present as evidence that EvoDuet applies beyond a single search framework.