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

Young-Jun Lee1 Jinheon Baek2 Soyeong Jeong2 Minki Kang2 Seungyeon Jwa1,3
Jonghyun Choi3 Seungho Han4 Dongyeop Kang1
1University of Minnesota2KAIST 3Seoul National University4Hanyang University
Paper Code Trajectories 7Best programs 11
An actual Swap Reduction run: retrieval at iteration 5 introduces exponential depth weighting, stored documents are reused at iteration 64, and retrieval at iteration 66 leads to a penalty for reversing the last SWAP. The final router uses 14,835 Q20 SWAPs.
EvoDuet co-evolves solutions and web searches. Q20 routing: 15,186 → 14,835 SWAPs. Run ↗

Abstract

EvoDuet co-evolves solutions and web search queries for scientific discovery, with fixed model parameters. A retrieval gate chooses between new searches, stored documents, and existing knowledge. Queries are refined using predicted solution scores; evaluated candidates inform later searches. Across 21 optimization tasks, EvoDuet improves discovery with GPT-5.6-Luna and Gemini-3.8-Flash, but not Qwen3.5-9B. Selected best runs improve on eight previously reported scores and match three.

Method

EvoDuet framework: an outer solution optimization loop connects through a knowledge-gap retrieval gate to an inner query optimization loop and a shared search database.
The outer loop evolves solutions; the inner loop refines searches. The gate selects Retrieve, Look-Up, or No-Op.

Discovery across 21 tasks

All scores ↗

EvoDuet improves discovery with GPT-5.6-Luna and Gemini-3.8-Flash. Qwen3.5-9B remains below the baseline.

With and without EvoDuet

Mean best Normalized Discovery Gain (NDG) ↑

GPT-5.6-Luna

+3.9 pp
OpenEvolve
74.1%
+ EvoDuet
78.0%

Gemini-3.8-Flash

+21.0 pp
OpenEvolve
61.3%
+ EvoDuet
82.3%

Qwen3.5-9B

−14.4 pp
OpenEvolve
66.8%
+ EvoDuet
52.4%

0% = initial program · 100% = reference · pp = percentage points

8 new best scores, 3 matches

All 11 best programs →
A circuit replay comparing the previous router with EvoDuet’s blue-and-purple circuit layout.
Quantum computing

Fewer SWAPs

15,18614,835

Q20 SWAP count ↓
The Rosetta encounter timeline, with the previous trajectory in purple and EvoDuet in blue.
Astrodynamics

A better Rosetta transfer

1.5529681.396424

Total Δv, km/s ↓
Single-cell gene expression before and after denoising, shown in the current paper’s blue-and-purple heatmaps.
Scientific algorithms

Cleaner single-cell counts

0.7226900.722906

Held-out denoising score ↑

Previous best → EvoDuet, from selected runs. Rosetta and Erdős refine published artifacts. Matches are at displayed precision.

Why the search works

Explore the trajectories →

Search design

GPT-5.6-Luna

Bi-level search outperforms joint and sequential search.

Native task scores, higher is better.
MethodMoleculeBurgersCircle
packing
OpenEvolveNo search0.84960.69372.635983
Joint-levelIn-loop search0.84740.69192.635980
DeepEvolveSequential0.81490.66662.581971
EvoDuetBi-level0.85240.78462.635983

Task scores ↑ · Circle packing: n = 26, tied at shown precision.

Searching for a knowledge gap beats fixed rules.

Normalized Discovery Gain, percent. Higher is better.
Retrieval gateDenoisingErdős
Random58.593.9
Heuristic60.699.5
Knowledge gap84.5100.0

NDG (%) ↑ · +23.9 pp over heuristic gating on Denoising.

The gains carry across three optimizers.

NDG gains in percentage points from adding EvoDuet.
OptimizerSums/DiffsDenoising
OpenEvolve+23.5+84.5
Top-K+46.7+63.6
EvoX+55.1+37.8

NDG gain (pp) from adding EvoDuet to each optimizer.

Methods from the literature

Runs with behavior (%)
Evidence-basedReuse / over-reach
55 of 82 runs transfer methods; 6 reuse published solutions. Behaviors can overlap.

6.9× lower cost on Denoising

EvoDuet reaches 100.27% NDG at $38.45. SimpleTES reaches 100% at an estimated API-equivalent $265.39.

View the denoising program →
Held-out NDG (%) ↑
Cost (USD) · log scale
Pareto frontier
Best observed results. SimpleTES cost is estimated; the bar shows 0.5–2× workload sensitivity.

Citation

BibTeX
Download .bib
Show BibTeX
@misc{lee2026evoduet,
  title  = {{EvoDuet}: Bilevel Co-Evolution of Web Searching
            and Task Solving for Scientific Discovery},
  author = {Lee, Young-Jun and Baek, Jinheon and Jeong, Soyeong
            and Kang, Minki and Jwa, Seungyeon and Choi, Jonghyun
            and Han, Seungho and Kang, Dongyeop},
  year   = {2026},
  eprint = {2609.40340},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url    = {https://arxiv.org/abs/2609.40340}
}

Paper figure