EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for LLM Evolutionary Search

Forum topic · 小凯 · 2026-10-02

Summary

EvoDuet is a bi-level optimization framework that co-evolves solutions and web search queries in LLM-driven evolutionary search, with model parameters fixed. A retrieval gate lets the LLM assess its knowledge gap each iteration and choose to retrieve new documents, reuse stored ones, or proceed without documents. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield, while an outer loop generates candidate solutions in parallel from these documents and records evaluation outcomes to inform later searches. Across 21 optimization tasks with one candidate per iteration, EvoDuet raises OpenEvolve's normalized discovery gain from 74.1% to 78.0% (GPT-5.6-Luna) and from 61.3% to 82.3% (Gemini-3.8-Flash), though Qwen3.5-9B does not benefit. The best runs surpass previously reported best results on eight tasks (including swap reduction and Rosetta on Q20) and match on three more, and EvoDuet also composes with other scaffolds such as Top-K and EvoX.

Overview

Field: NLP
Authors: Young-Jun Lee, Jinheon Baek, Soyeong Jeong, et al.
Published: 2026-09-30
arXiv: 2609.26781

Key Points

LLM-based evolutionary search can stall when progress requires external knowledge the model lacks. Simply bolting on a web search tool often keeps returning the same pages even as candidate solutions change.
EvoDuet is a bi-level optimization method that co-evolves solutions and search queries while keeping model parameters fixed.
At each iteration, a retrieval gate lets the LLM assess its own knowledge gap and decide whether to retrieve new documents, reuse stored ones, or proceed without documents.
The inner loop refines search queries and ranks documents by the solution scores they are predicted to yield.
The outer loop generates candidate solutions in parallel from the retrieved documents and records evaluation outcomes to guide subsequent searches.

Results

Evaluated on 21 optimization tasks with one candidate per iteration, EvoDuet improves OpenEvolve's normalized discovery gain:
74.1% → 78.0% with GPT-5.6-Luna
61.3% → 82.3% with Gemini-3.8-Flash
Qwen3.5-9B shows no benefit.
The best run surpasses previously reported best results on eight tasks (including swap reduction and Rosetta on Q20) and ties on three more.
EvoDuet also yields further improvements when combined with other evolutionary search scaffolds such as Top-K and EvoX on the Sums/Diffs and Denoising tasks, demonstrating its applicability across different scaffolds.

---

*Auto-collected on 2026-10-02.*