Paper Overview
Field: NLP Authors: Young-Jun Lee, Jinheon Baek, Soyeong Jeong, et al. Published: 2026-09-30 arXiv: 2609.26781
Summary
Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding a web search tool can keep returning the same pages as solutions change. EvoDuet is a bi-level optimization method that co-evolves solutions and search queries while keeping model parameters fixed.
How it works
- Retrieval gate: At each iteration, the LLM assesses its knowledge gap and chooses to retrieve new documents, reuse stored ones, or proceed without documents.
- Inner loop: Refines search queries and ranks documents by the solution scores they are predicted to yield.
- Outer loop: Generates candidate solutions in parallel from these documents and records evaluated outcomes to inform later searches.
- Evaluated on 21 optimization tasks with one candidate per iteration.
- EvoDuet improved OpenEvolve's normalized discovery gain from 74.1% to 78.0% using GPT-5.6-Luna, and from 61.3% to 82.3% using Gemini-3.8-Flash.
- Qwen3.5-9B did not benefit from the method.
- The best run surpassed previously reported best results on eight tasks (including swap reduction on Q20 and Rosetta) and matched them on three additional tasks.
- EvoDuet combines well with other scaffolds such as Top-K and EvoX, delivering further improvements on Sums/Diffs and Denoising tasks, demonstrating applicability across different evolutionary search scaffolds.
Results
*Auto-collected on 2026-10-02*