acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions change. We introduce EvoDuet, a bi-level optimization method that co-evolves solutions and search queries with fixed model parameters. At each iteration, a retrieval gate lets the LLM assess its knowledge gap and choose to retrieve new documents, reuse stored ones, or proceed without them. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield; an outer loop generates candidates in parallel from these documents and records the evaluated outcomes for later searches. Across 21 optimization tasks with one candidate per iteration, EvoDuet raises 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, whereas Qwen3.5-9B does not benefit. Our best runs surpass the previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three more. EvoDuet also improves with other scaffolds (e.g., Top-K, EvoX) on Sums/Diffs and Denoising, demonstrating its applicability across evolutionary search scaffolds.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
scientific discovery, bi-level optimization, web searching, co-evolution

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_949aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EvoDuet integrates web retrieval into evaluator-driven evolutionary program search. It enables optimizers to proactively acquire external knowledge using a combination of knowledge-gap gating, iterative query refinement, search memory, and utility-driven evidence selection based on predicted downstream evaluator scores. The core research direction is important, and the proposed system demonstrates compelling empirical potential.

Critical strengths
  1. Clean writing and easy to understand.
  2. Selecting evidence based on predicted downstream task utility rather than raw semantic relevance is a well-motivated and promising design principle.
  3. Evaluation: The empirical evaluation spans 21 main optimization tasks across multiple models and transparently reports negative results. The method achieves substantial performance improvements on notable tasks such as Swap Reduction.
Critical weaknesses
  1. Overclaim: The theoretical claim of bilevel optimization is overstated. The actual inner loop relies on a frozen LLM acting as a self-predicted heuristic surrogate without formal optimization guarantees. Labeling the query loop as "co-evolution" is strong. The mechanism operates primarily as iterative query refinement or active retrieval.
  2. Compute not matched: EvoDuet adds up to RJ+R+2 LLM calls and up to three web searches per RETRIEVE iteration, but baselines are matched only on candidate count. Cost is reported only for Table 1a and Denoising. A baseline with matched tokens or LLM calls (e.g., more samples, self-reflection) is needed, especially since parallel generation alone helps OpenEvolve a lot (Appendix D.2).
  3. The key comparison against DeepEvolve is not convincing enough.
  4. Combining open public benchmarks with unrestricted web search blurs the line between genuine "scientific discovery" and literature-assisted solution reconstruction.
Questions for authors
  1. Do you have stronger compute-matched baselines, e.g., OpenEvolve with more sampling/self-reflection? It would also be useful to report score vs. tokens / LLM calls / evaluator calls.
  2. Do you have ablations for the utility scorer, e.g., comparing it with semantic relevance ranking, standard reranking, or random document selection?
  3. Do you have any evaluation on tasks where the target solution cannot be directly found or reconstructed from the public web, even on a small subset?
4 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

W1. Overclaim: The theoretical claim of bilevel optimization is overstated. The actual inner loop relies on a frozen LLM acting as a self-predicted heuristic surrogate without formal optimization guarantees. Labeling the query loop as "co-evolution" is strong. The mechanism operates primarily as iterative query refinement or active retrieval.

A. Thank you for highlighting this issue. We agree that describing our method as “bilevel optimization” may overstate its theoretical scope, since the inner loop relies on an LLM-predicted surrogate objective without formal optimization guarantees. We used this formulation to emphasize the mutual influence between the search and solution loops, drawing on prior work on surrogate-assisted bilevel optimization [1,2]. However, we agree that the term “bilevel” should be used more carefully. We will therefore revise the manuscript to use “co-evolution” instead, referring specifically to the interaction between query refinement and solution optimization.


(Q1 & W2). Stronger baselines under matched token budget

A. We agree that matching candidate counts alone does not account for retrieval overhead. We added OpenEvolve baselines with more sampling and self-reflection, using the same recorded LLM token caps as EvoDuet n=1 (GPT-5.6-Luna).

TaskMethodScore ↑Tokens (M)LLM callsEvaluator callsSearch calls
DenoisingEvoDuet n=10.715211.84641744116
DenoisingOpenEvolve + more sampling0.711412.0481,1161,0870
DenoisingOpenEvolve + self-reflection0.671212.0751,5239750
Sums/DiffsEvoDuet n=11.09861017.6003983897
Sums/DiffsOpenEvolve + more sampling1.08037217.6036055410
Sums/DiffsOpenEvolve + self-reflection1.07725717.6687554950

The table reports higher scores for EvoDuet than for OpenEvolve with more sampling or self-reflection, alongside similar token usage and fewer LLM calls.


W3. The key comparison against DeepEvolve is not convincing enough.

A. We expanded the DeepEvolve comparison to all 21 tasks using GPT-5.6-Luna. For robustness, we run three times.

DomainOpenEvolveDeepEvolveEvoDuet
Mathematics (8)71.0175.7475.50
Astrodynamics (5)78.4277.2580.66
Scaling laws (4)53.2154.4552.37
Denoising (1)0.0036.8474.39
Swap reduction (1)21.8335.8843.76
AHC (2)46.0739.2746.08
Overall (21)61.2864.8267.96

In the Table, EvoDuet has the highest overall point estimate (+3.14 NDG points over DeepEvolve), with larger gains in Denoising and Swap. DeepEvolve remains competitive in mathematics and scaling laws. We will add the full results to the revised paper.


W4. Combining open public benchmarks with unrestricted web search blurs the line between genuine "scientific discovery" and literature-assisted solution reconstruction.

A. Thank you for raising this concern. The table below reports SOTA matches and improvements by observed behavior.

Observed behaviorMatches SOTAExceeds SOTATotal SOTA results
Method transfer088
Cross-domain transfer022
Live documentation lookup123
Target-informed reconstruction347
Public artifact reuse022
Evidence-inspired over-reach033
All reported results (unique)3811

Although some SOTA results involve public artifact reuse or target-informed reconstruction, method transfer also appears in eight runs that exceed SOTA. Appendix I provides concrete examples of how models apply principles from the literature to revise their code, supporting the value of web search for literature-assisted optimization. However, we agree that such optimization should be distinguished from scientific discovery and will clarify this point in the revised manuscript.

AuthorsOpenReview ↗

Response by Authors (2/2)

Q2. Do you have ablations for the utility scorer, e.g., comparing it with semantic relevance ranking, standard reranking, or random document selection?

A. Thank you for this suggestion. We additionally evaluated random selection, semantic relevance ranking, and standard reranking with BGE on Denoising, Sums/Diffs, and Swap Reduction using GPT-5.6-Luna.

Document selectionDenoising (PBMC/Tabula) ↑Sums/Diffs (c-value) ↑Swap (Q20 SWAPs) ↓
Random selection0.6744 ± 0.04021.081573 ± 0.00691918,611 ± 569
Semantic relevance0.6276 ± 0.02931.111068 ± 0.01588119,113 ± 957
BGE reranker0.6901 ± 0.02621.079084 ± 0.01043919,841 ± 226
EvoDuet0.7019 ± 0.02011.087436 ± 0.01064917,894 ± 1,579

As shown in the table, EvoDuet’s hypothetical evidence scoring achieves better mean performance than the alternative selection methods, except on Sums/Diffs, where semantic relevance ranking performs best.


Q3. Do you have any evaluation on tasks where the target solution cannot be directly found or reconstructed from the public web, even on a small subset?

A. We conduct an additional experiment on ConStellaration [3], a dataset and machine learning optimization benchmark for quasi-isodynamic (QI) stellarator plasma boundary shapes. No public solution is currently available. For robustness, we run the experiment three times.

MethodConStellaration score ↑
OpenEvolve0.432046
EvoDuet (n=1)0.435753
EvoDuet (n=8)0.435825
EvoDuet (n=16)0.436631

As shown in the table, EvoDuet outperforms OpenEvolve and exhibits method transfer. This involves exploring principles in the scientific literature and adapting them into executable code, including Fourier-mode regularization and evaluator-guided coefficient tuning. These results suggest that co-evolution between web search and solution optimization is helpful even on tasks where no public solution exists.


References:

[1] Dumouchelle, Justin, et al. "Neur2bilo: Neural bilevel optimization." Advances in Neural Information Processing Systems 37 (2024): 86688-86719.

[2] Chew, Ruth Wan Theng, Quoc Phong Nguyen, and Bryan Kian Hsiang Low. "BILBO: BILevel Bayesian Optimization." arXiv preprint arXiv:2502.02121 (2025).

[3] Cadena, Santiago, et al. "ConStellaration: A dataset of QI-like stellarator plasma boundaries and optimization benchmarks." Advances in Neural Information Processing Systems 38 (2026).

Thank you for the additional experiments. They address parts of my concerns.

Considering the contribution is primarily integrating established components, such integration can be valuable, but it would need stronger evidence than the current ablations and the modest gain over DeepEvolve provide. For example, denoising accounts for ~half of the reported average gain over OpenEvolve. But token-matched OpenEvolve reaches 0.7114 on denoising, better than the average of EvoDuet and achieves ~86% NDG. Given that standard OpenEvolve scored 0 NDG in the 21-task comparison, and the budget-matched comparisons cover only 2 tasks. I remain unsure how much of EvoDuet’s reported advantage persists under matched budgets. I am therefore maintaining my rating of 2.

AuthorsOpenReview ↗

Response by Authors

Thank you for clarifying your concern. We agree that two tasks are insufficient to establish a general advantage under matched token budgets. We will extend the token-matched evaluation to the remaining 19 tasks to provide broader evidence, when we acquire sufficient cost (e.g., API cost) for the experiment.

Reviewer_UxSfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoDuet framework that combines program optimization with iterative web search. The setup is for programming tasks with search. The key module is the process of information searching. The inner loop refines search queries and outer loop generates and evaluates candidate programs.

Critical strengths
  1. The paper is clearly written. The figures and examples help explain how the two loops work together. 2.The discussion is clear and useful. The authors show how retrieved information affects program changes.
Critical weaknesses

It is unclear whether separating search and code generation into two loops is better than letting an agent flexibly decide when to perform each action (ReAct agent, or codex, etc.). Clearer explanation or a comparison would be better for me to understand.

Questions for authors
  1. I am wondering if the proposed framework is intended only for programming tasks with a numerical evaluator. For deep research tasks where the model already knows what to search for, would the inner loop still help? How should the number of search rounds be chosen for such tasks?
3 comments
AuthorsOpenReview ↗

Response by Authors

W1. It is unclear whether separating search and code generation into two loops is better than letting an agent flexibly decide when to perform each action (ReAct agent, or codex, etc.). Clearer explanation or a comparison would be better for me to understand.

A. We evaluate ReAct agent, Codex (w/ GPT-6.1-Sol), and Claude-Code (w/ Opus 5.5) on 21 tasks. For robustness, we run each agent with three different random seeds.

TaskReAct (GPT-5.6-Luna)Codex (GPT-6.1-Sol)Claude Code (Opus 5.5)OpenEvolve (GPT-5.6-Luna)EvoDuet (n=1, GPT-5.6-Luna)EvoDuet (n=8, GPT-5.6-Luna)
Erdos52.7%100.0%99.8%95.8%99.9%98.5%
AC111.0%105.7%59.0%42.8%41.2%64.8%
AC220.8%94.6%48.6%41.2%65.9%72.1%
AC396.9%99.9%99.6%99.7%99.5%99.8%
CP (n=26)49.5%100.0%100.0%100.0%100.0%100.0%
CP (n=32)93.4%100.0%99.9%100.0%99.8%100.0%
Hadamard37.8%100.0%95.4%73.0%65.3%98.6%
Sums/Diffs6.6%100.8%78.8%15.4%32.5%53.8%
Cassini0.0%100.0%99.0%73.8%87.2%82.7%
Galileo0.8%103.3%102.6%37.6%37.7%47.2%
Mariner 1028.8%100.0%100.0%100.0%100.0%100.0%
Rosetta15.2%101.8%100.9%81.4%79.7%86.8%
Voyager 20.0%100.0%100.0%99.3%98.7%99.3%
Domain Mixture Scaling84.4%98.8%0.0%97.5%97.6%98.6%
Easy Question Scaling0.0%0.5%0.0%0.0%0.0%0.0%
LR/BSZ Scaling38.8%82.6%76.7%23.9%28.3%35.7%
Parallel Scaling34.5%98.4%85.4%91.4%83.6%99.1%
Denoising0.0%97.0%94.1%0.0%74.4%71.4%
Swap Reduction5.4%260.1%65.4%21.8%43.8%48.0%
AHC0390.0%0.0%52.7%0.0%0.0%0.0%
AHC05840.1%83.3%88.9%92.1%92.2%87.2%
Avg29.4%96.5%78.4%61.3%68.0%73.5%

When compared with stronger frontier agents such as Codex and Claude Code, our method achieves a lower average NDG. However, on AHC058, EvoDuet achieves a higher NDG than both frontier agents, despite their better overall performance. When compared with the ReAct agent, which is relatively weaker than the frontier agents, EvoDuet achieves better average performance (68.0% vs. 29.4% NDG), suggesting that the two-loop co-evolution approach can be effective in this setting. Thank you for suggesting this comparison with other agents. We have revised the PDF to include this experiment.


Q1. I am wondering if the proposed framework is intended only for programming tasks with a numerical evaluator. For deep research tasks where the model already knows what to search for, would the inner loop still help? How should the number of search rounds be chosen for such tasks?

A. Yes, we initially focused on scientific optimization tasks with deterministically verifiable numerical evaluators. However, based on your comment, we can extend EvoDuet to deep research tasks by changing the evaluator component. Hence, we additionally evaluated OpenEvolve and EvoDuet on DeepResearch-Bench-II [1], which requires long-form answers. We varied the number of search rounds (1, 3, and 5), following the setting in Table 6(a) of Appendix D.3. Scores are the mean percentage of benchmark rubric criteria met. Model: GPT-5.6-Luna.

MethodSearch roundsDRB-II score (%) ↑
OpenEvolve—28.9
EvoDuet128.1
EvoDuet327.8
EvoDuet531.7

We observe that EvoDuet can also outperform OpenEvolve on this deep research task. However, increasing the number of search rounds from 1 to 3 slightly decreases performance; five rounds yield the highest score (31.7%, compared with 28.9% for OpenEvolve).


References

[1] Li, Ruizhe, et al. "Deepresearch bench ii: Diagnosing deep research agents via rubrics from expert report." arXiv preprint arXiv:2601.08536 (2026).

Thank you for the additional experiments. They help, but I still have two concerns.

  1. The comparison with ReAct is informative. It looks like these optimization tasks are challenging for simple ReAct. But why was the codex tested on a stronger model (6.1 sol)? It makes an unfair comparison. Based on the results alone, it seems like the general framework codex can handle these tasks very well, outperforming your baselines and proposed framework.

  2. The new results show that the effect of search rounds depends on the task, but it is still unclear how users should choose this number when applied to new tasks. On Denoising, increasing the rounds from 3 to 5 lowers NDG from 81.77% to 34.28%, whereas five rounds perform best among the tested settings on DRB-II. Is there a signal for stopping search when further rounds are unlikely to help?

I will raise my score if these concerns are addressed.

AuthorsOpenReview ↗

Response by Authors

Q1. The comparison with ReAct is informative. It looks like these optimization tasks are challenging for simple ReAct. But why was the codex tested on a stronger model (6.1 sol)? It makes an unfair comparison. Based on the results alone, it seems like the general framework codex can handle these tasks very well, outperforming your baselines and proposed framework.

A. We additionally evaluate Codex (w/ GPT-5.6-Luna), the same backbone used by EvoDuet, on all 21 tasks. For robustness, we run three times.

TaskCodex (GPT-5.6-Luna)Codex (GPT-6.1-Sol)EvoDuet (n=1, GPT-5.6-Luna)EvoDuet (n=8, GPT-5.6-Luna)
Erdos54.1%100.0%99.9%98.5%
AC115.8%105.7%41.2%64.8%
AC262.4%94.6%65.9%72.1%
AC398.8%99.9%99.5%99.8%
CP (n=26)99.8%100.0%100.0%100.0%
CP (n=32)99.1%100.0%99.8%100.0%
Hadamard51.4%100.0%65.3%98.6%
Sums/Diffs25.9%100.8%32.5%53.8%
Cassini28.4%100.0%87.2%82.7%
Galileo28.4%103.3%37.7%47.2%
Mariner 1061.4%100.0%100.0%100.0%
Rosetta41.3%101.8%79.7%86.8%
Voyager 253.7%100.0%98.7%99.3%
Domain Mixture Scaling97.0%98.8%97.6%98.6%
Easy Question Scaling0.0%0.5%0.0%0.0%
LR/BSZ Scaling49.9%82.6%28.3%35.7%
Parallel Scaling56.6%98.4%83.6%99.1%
Denoising17.8%97.0%74.4%71.4%
Swap Reduction7.1%260.1%43.8%48.0%
AHC0390.8%0.0%0.0%0.0%
AHC05890.7%83.3%92.2%87.2%
Avg49.6%96.5%68.0%73.5%

In the Table, EvoDuet achieves a higher average NDG using the same backbone.


Q2. The new results show that the effect of search rounds depends on the task, but it is still unclear how users should choose this number when applied to new tasks. On Denoising, increasing the rounds from 3 to 5 lowers NDG from 81.77% to 34.28%, whereas five rounds perform best among the tested settings on DRB-II. Is there a signal for stopping search when further rounds are unlikely to help?

A. Thank you for this helpful comment. We agree that the fixed search budget is a limitation of the current proposed framework. Once retrieval is invoked by knowledge-gap-based gating, EvoDuet completes the predefined number of inner search rounds without early stopping, even when additional rounds provide little useful evidence. A potential stopping signal is that successive rounds neither improve hypothetical evidence scores nor change the retained top-D document set, although we have not yet validated this criterion.


We hope our response addresses your concerns. Please let us know if you have any further questions or comments.

Reviewer_NXtcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LLM-driven evolutionary search stalls when progress requires knowledge the model does not have, and simply attaching a web search tool leads the model to retrieve the same pages over and over. EvoDuet addresses this by co-evolving solutions and search queries in a bi-level loop with fixed model parameters. A knowledge-gap gate lets the model decide at each iteration whether to retrieve new documents, reuse stored ones, or continue without them. When it retrieves, an inner loop refines queries and ranks documents by the evaluator score each one is predicted to yield, without generating or evaluating any candidate. The outer loop then generates several candidates in parallel from the selected documents and records the outcomes to guide later searches. Across a suite of optimization tasks, EvoDuet improves discovery over the base scaffold for two strong frontier models, though a smaller open model does not benefit. Best runs surpass or match previously reported scores on a number of tasks, the method reaches state-of-the-art quality at lower cost on one task, and it transfers to other evolutionary scaffolds.

Critical strengths
  1. The motivation is well grounded, and the analysis is informative.
  2. The evaluation is broad and produces real results.
Critical weaknesses
  1. The novelty appears primarily integrative, combining adaptive retrieval, query refinement, and document selection within evolutionary search. The paper does not sufficiently clarify what distinct methodological contribution the bi-level design offers beyond this integration.
  2. The benchmark rewards retrieval rather than discovery. NDG is normalized to the best-known human result, and since the core mechanism is web search, a perfect score can mean the answer was simply found online. Gains concentrate exactly where the optimum is publicly posted, and no stratified check by retrievability is provided.
Questions for authors
  1. Best-of-four results may obscure typical performance. Please emphasize mean gains, confidence intervals, and task-level win rates.
  2. Gate selection and candidate counts confound retrieval benefits. Please conduct paired comparisons with identical parents and candidate budgets.
  3. Please stratify results by solution retrievability and distinguish between retrieving known solutions, improving them, and producing new discoveries.
4 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

W1. The novelty appears primarily integrative, combining adaptive retrieval, query refinement, and document selection within evolutionary search. The paper does not sufficiently clarify what distinct methodological contribution the bi-level design offers beyond this integration.

A. Thank you for the comment. Our contribution is to guide retrieval by its predicted effect on the current solution. The inner loop holds the solution fixed while refining queries and selecting documents using predicted solution scores. The outer loop then generates and evaluates a new solution, and the measured outcome informs subsequent retrieval decisions.

We compare joint-level and two-level search in (a) and report component ablations in (b).

(a) Joint-level vs. bi-level search (Appendix B.5, Table 2)

Search designMean NDG ↑
Joint-level63.10
EvoDuet (bi-level)81.49

(b) Additional component ablations (Reviewer 7rG2’s W3 & Q4)

We ablated four components across four tasks using GPT-5.6-Luna. The ablations share a frozen document corpus per task to control retrieval variation.

Component ablationDenoising ↑Sums/Diffs ↑Voyager 2 ↓Swap ↓
Full EvoDuet0.70191.0874363.438317,894
w/o knowledge-gap gating0.70971.0864653.44638719,343
w/o query refinement0.70791.0864303.43020617,174
w/o hypothetical evidence scoring0.56871.0734303.44638719,719
w/o persistent search memory0.70001.0864713.44641119,260

The results in (a) show that bi-level search outperforms joint-level search. Among the ablations in (b), we observe a drop in performance when a key component is removed.


W2. The benchmark rewards retrieval rather than discovery. NDG is normalized to the best-known human result, and since the core mechanism is web search, a perfect score can mean the answer was simply found online. Gains concentrate exactly where the optimum is publicly posted, and no stratified check by retrievability is provided.

A. Thank you for raising this concern. We agree that NDG measures progress relative to a reference score and does not by itself establish scientific discovery. The table below reports SOTA matches and improvements by observed behavior.

Observed behaviorMatches SOTAExceeds SOTATotal SOTA results
Method transfer088
Cross-domain transfer022
Live documentation lookup123
Target-informed reconstruction347
Public artifact reuse022
Evidence-inspired over-reach033
All reported results (unique)3811

Although some SOTA results involve public artifact reuse or target-informed reconstruction, method transfer also appears in eight runs that exceed SOTA. Appendix I provides concrete examples of how models apply principles from the literature to revise their code. To examine performance when no public solution is available, we additionally evaluated EvoDuet on ConStellaration [1], a benchmark for optimizing quasi-isodynamic (QI) stellarator plasma boundary shapes. We ran the experiment three times for robustness.

MethodConStellaration score ↑
OpenEvolve0.432046
EvoDuet (n=1)0.435753
EvoDuet (n=8)0.435825
EvoDuet (n=16)0.436631

As shown in the table, EvoDuet achieves higher scores than OpenEvolve. The runs also exhibit method transfer through code revisions such as Fourier-mode regularization and evaluator-guided coefficient tuning. Together, the behavior analysis and the ConStellaration results suggest that EvoDuet’s gains are not explained solely by retrieving publicly available solutions. We will clarify the distinction between literature-assisted optimization and scientific discovery in the revised manuscript.


References:

[1] Cadena, Santiago, et al. "ConStellaration: A dataset of QI-like stellarator plasma boundaries and optimization benchmarks." Advances in Neural Information Processing Systems 38 (2026).

AuthorsOpenReview ↗

Response by Authors (2/2)

Q1. Best-of-four results may obscure typical performance. Please emphasize mean gains, confidence intervals, and task-level win rates.

A. We already report per-task mean ± standard deviation in Appendix D.4. We summarize mean NDG across all 21 tasks below and additionally calculate task-level win rates and exploratory confidence intervals.

BackboneNOpenEvolve mean NDGEvoDuet mean NDGMean gain [95% CI]Task W/T/L (win rate)
Qwen3.5-9B143.6342.79-0.84 [-8.12, 6.99]8/0/13 (38.1%)
Qwen3.5-9B849.7151.21+1.49 [-9.38, 11.72]12/0/9 (57.1%)
GPT-5.6-Luna162.3867.95+5.56 [0.47, 11.86]9/4/8 (42.9%)
GPT-5.6-Luna863.7773.52+9.75 [1.70, 17.35]14/0/7 (66.7%)
Gemini-3.8-Flash154.7670.33+15.57 [5.48, 26.16]14/4/3 (66.7%)
Gemini-3.8-Flash882.5876.82-5.76 [-15.40, 2.15]7/5/9 (33.3%)

EvoDuet improves mean NDG with GPT-5.6-Luna at both N=1 and N=8, and with Gemini-3.8-Flash at N=1. The gains vary across backbones and sampling settings.

ModelMethodNQuantum compilation (1)Astrodynamics (5)Scientific algorithms (1)AI foundations (4)Algorithm engineering (2)Mathematics (8)Overall (21)
Qwen3.5-9BOpenEvolve15.7747.4117.8644.4930.0352.1943.63
Qwen3.5-9BEvoDuet10.0045.2820.1931.2333.1657.6042.79
Qwen3.5-9BOpenEvolve85.8363.6478.4850.2946.0443.5349.71
Qwen3.5-9BEvoDuet86.3157.2541.4643.6544.1059.8251.21
GPT-5.6-LunaOpenEvolve121.8878.4222.0953.2046.7470.9662.38
GPT-5.6-LunaEvoDuet143.7580.6674.3552.3846.0875.4867.95
GPT-5.6-LunaOpenEvolve840.8973.9060.5161.5248.6865.6063.77
GPT-5.6-LunaEvoDuet848.0083.2271.4558.3443.6285.9773.52
Gemini-3.8-FlashOpenEvolve156.6430.0085.8545.7147.0872.5654.76
Gemini-3.8-FlashEvoDuet124.3454.9884.4673.6448.1387.8070.33
Gemini-3.8-FlashOpenEvolve854.79100.8184.4062.4871.6487.2282.58
Gemini-3.8-FlashEvoDuet859.6173.4061.5261.8467.6292.8176.82

Q2. Gate selection and candidate counts confound retrieval benefits. Please conduct paired comparisons with identical parents and candidate budgets.

A. We conducted paired comparisons on Denoising, Sums/Diffs, Voyager 2, and Swap Reduction with GPT-5.6-Luna. For each of three saved parent snapshots, we planned four candidate generations without documents and four with three documents selected by hypothetical evidence scoring. Both conditions used the same parent code and base prompt, with identical candidate budgets.

TaskWithout documentsWith documentsΔEvaluated pairs
Denoising0.16520.2396+0.074412/12
Sums/Diffs1.06381.0694+0.005612/12
Voyager 20.66180.4983-0.16358/12
Swap Reduction4,677.96,466.2+1,788.37/12
Retrieved documents improved mean candidate scores in three of four tasks.

Q3. Please stratify results by solution retrievability and distinguish between retrieving known solutions, improving them, and producing new discoveries.

A. Thank you for this suggestion. We extend the behavior analysis in Figure 8(c) and Appendix E.1 with score changes and concrete examples.

Observed behaviorRuns / 82 (%)Median ΔNDG
Method transfer55 (67.1%)+1.24
Cross-domain transfer13 (15.9%)+1.54
Live documentation lookup27 (32.9%)+0.05
Target-informed reconstruction40 (48.8%)+4.97
Public artifact reuse6 (7.3%)0.00
Evidence-inspired over-reach29 (35.4%)−16.08
Case / source useWhat EvoDuet didRecorded result
Public artifact reuse — Erdős (§I.5)Downloaded and validated a published vector.C₅: 0.381017 → 0.380859 ↓
Method transfer — Swap (§I.1)Implemented and tuned dSABRE depth weighting.Q20 SWAPs: 19,953 → 18,030; final 14,835 ↓
Cross-domain transfer — Erdős (§I.2)Adapted and refined a radar codeword for a number-theory construction.C₅: 0.381374 → 0.380915 ↓
Target-informed reconstruction — CP, n=26 (§I.4)Built an LP + SLSQP solver using a known target, without retrieved coordinates.Sum of radii: 1.390 → 2.635983 (reference match) ↑

These examples distinguish direct reuse of a known solution from reconstructing or improving solutions using retrieved knowledge. In the revised manuscript, we will clarify what was retrieved and what the model contributed, while making clear that improved benchmark scores alone do not establish scientific discovery.

Thank you for addressing all my concerns. I will raise my rating to 3.

AuthorsOpenReview ↗

Response by Authors

We are glad that our response, including the additional experiments, has addressed your concerns, and we thank you for raising your score. We will incorporate these experiments and clarifications into the revised paper.

Reviewer_7rg2OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoDuet, a bilevel framework that couples web search with evolutionary solution optimization. The outer loop evolves candidate solutions, while the inner loop generates and refines search queries based on the model’s estimated knowledge gap. Retrieved documents are ranked using a hypothetical evidence score that predicts their downstream utility for improving the evaluator score, and selected documents are then used to generate new candidates. The method is evaluated across a broad collection of scientific and algorithmic optimization tasks and shows improvements over OpenEvolve on several model backbones and task groups.

The overall idea is interesting and the empirical study is broad. However, I have concerns about whether the current experiments sufficiently distinguish genuine scientific discovery from retrieval-assisted optimization, and whether the reported gains robustly isolate the contribution of the proposed bilevel mechanism.

Critical strengths

The paper addresses an important limitation of existing LLM-based optimization systems: they typically rely on fixed parametric knowledge or previously observed trajectories, while EvoDuet allows the agent to actively acquire external information during optimization.

The idea of optimizing retrieval according to downstream evaluator performance is interesting. In particular, hypothetical evidence scoring is more closely aligned with the optimization objective than standard relevance-based retrieval.

The evaluation covers a diverse set of domains, including scientific optimization, quantum compilation, astrodynamics, mathematical constructions, and algorithm engineering, which provides useful evidence about when external retrieval can help.

The paper includes several useful analyses of retrieval behavior, prediction quality, document usage, and failure cases. I also appreciate that negative results, such as weaker performance on some backbones and algorithm-engineering tasks, are reported explicitly.

Critical weaknesses
  1. The evaluation does not sufficiently distinguish scientific discovery from retrieval and adaptation of existing public solutions. Several tasks have publicly available trajectories, constructions, parameters, implementations, or known solutions. In such cases, improved performance may come from finding and adapting existing artifacts rather than discovering genuinely new solutions. The analysis of direct code or numeric copying is useful, but it does not establish conceptual novelty.

  2. The main evaluation reports the best result across four runs rather than the mean. This is particularly concerning for a retrieval-based method because EvoDuet introduces substantial additional variance through query generation, web search, document selection, and candidate generation. A strong best run may partly reflect an unusually favorable retrieval trajectory rather than a consistently better optimization method.

  3. The contribution of the proposed bilevel query optimization is not cleanly isolated from retrieval randomness and other system components. EvoDuet simultaneously introduces knowledge-gap gating, iterative query refinement, hypothetical evidence scoring, persistent search memory, and additional search/model calls. The current ablations are limited to a relatively small subset of tasks, making it difficult to identify which component is responsible for the overall gains.

  4. The validation of hypothetical evidence scoring is useful but does not directly measure the causal value of individual documents. The model predicts the evaluator score that would result from using one document, while the actual candidate may be generated from multiple retained documents jointly. Therefore, the reported correlation demonstrates that the surrogate is informative, but does not fully establish accurate prediction of each document’s downstream contribution.

Questions for authors
  1. Could the authors report mean ± standard deviation across the four runs for the 21-task main evaluation, rather than only the best run, and clarify whether the improvements remain statistically significant?

  2. How much of the gain comes from the proposed query optimization rather than retrieval randomness? A controlled experiment using cached or fixed search results would help isolate the contribution of the search policy.

  3. For the claimed discoveries or SOTA results, can the authors distinguish what was directly retrieved from public sources, what was adapted or recombined by the model, and what constitutes genuinely novel solution components?

  4. Could the authors provide a broader compute-matched ablation separating the effects of retrieval gating, iterative query evolution, hypothetical evidence scoring, and persistent document memory?

5 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

W1 & Q3. Distinguishing retrieval, adaptation, and discovery

A. Thank you for raising these related points. We agree that solution reuse, adaptation, and scientific discovery should be distinguished. We extend the behavior analysis in Figure 8(c) and trace retrieved sources to the model’s implementation and refinement.

Observed behaviorRuns / 82 (%)Median ΔNDG
Method transfer55 (67.1%)+1.24
Cross-domain transfer13 (15.9%)+1.54
Live documentation lookup27 (32.9%)+0.05
Target-informed reconstruction40 (48.8%)+4.97
Public artifact reuse6 (7.3%)0.00
Evidence-inspired over-reach29 (35.4%)−16.08
Observed behaviorMatches SOTAExceeds SOTATotal SOTA results
Method transfer088
Cross-domain transfer022
Live documentation lookup123
Target-informed reconstruction347
Public artifact reuse022
Evidence-inspired over-reach033
All reported results (unique)3811

The SOTA counts cover 11 selected task-level results across backbones and candidate budgets. Representative examples from Appendix I:

Task / appendixRetrieved contentModel’s contribution
Swap Reduction (§I.1)The dSABRE depth-weighting principle.Implemented and tuned it in the Rust router; final Q20 SWAPs: 14,835.
CP, n=26 (§I.4)A published target value, without solution coordinates.Built an LP + SLSQP solver; sum of radii: 2.635983 (reference match).
Erdős (§I.5)A published solution vector.Validated and refined the vector; approximately 2.6 × 10⁻¹⁰ improvement beyond the retrieved vector.

These analyses identify the components implemented, reconstructed, or refined by the model. We will distinguish them from directly retrieved content and clarify the scope of novelty supported by each case in the revised manuscript.


W2. The main evaluation reports the best result across four runs rather than the mean. This is particularly concerning for a retrieval-based method because EvoDuet introduces substantial additional variance through query generation, web search, document selection, and candidate generation. A strong best run may partly reflect an unusually favorable retrieval trajectory rather than a consistently better optimization method.

A. We already report the mean and standard deviation in Appendix D.4. We chose to report the best result because, in scientific discovery tasks, achieving state-of-the-art performance is often a primary goal. Below, we summarize overall performance based on mean scores and additionally calculate task-level win rates and exploratory confidence intervals.

BackboneNOpenEvolve mean NDGEvoDuet mean NDGMean gain [95% CI]Task W/T/L (win rate)
Qwen3.5-9B143.6342.79-0.84 [-8.12, 6.99]8/0/13 (38.1%)
Qwen3.5-9B849.7151.21+1.49 [-9.38, 11.72]12/0/9 (57.1%)
GPT-5.6-Luna162.3867.95+5.56 [0.47, 11.86]9/4/8 (42.9%)
GPT-5.6-Luna863.7773.52+9.75 [1.70, 17.35]14/0/7 (66.7%)
Gemini-3.8-Flash154.7670.33+15.57 [5.48, 26.16]14/4/3 (66.7%)
Gemini-3.8-Flash882.5876.82-5.76 [-15.40, 2.15]7/5/9 (33.3%)

EvoDuet improves mean NDG with GPT-5.6-Luna at both N=1 and N=8, and with Gemini-3.8-Flash at N=1. The gains vary across backbones and sampling settings.

ModelMethodNQuantum compilation (1)Astrodynamics (5)Scientific algorithms (1)AI foundations (4)Algorithm engineering (2)Mathematics (8)Overall (21)
Qwen3.5-9BOpenEvolve15.7747.4117.8644.4930.0352.1943.63
Qwen3.5-9BEvoDuet10.0045.2820.1931.2333.1657.6042.79
Qwen3.5-9BOpenEvolve85.8363.6478.4850.2946.0443.5349.71
Qwen3.5-9BEvoDuet86.3157.2541.4643.6544.1059.8251.21
GPT-5.6-LunaOpenEvolve121.8878.4222.0953.2046.7470.9662.38
GPT-5.6-LunaEvoDuet143.7580.6674.3552.3846.0875.4867.95
GPT-5.6-LunaOpenEvolve840.8973.9060.5161.5248.6865.6063.77
GPT-5.6-LunaEvoDuet848.0083.2271.4558.3443.6285.9773.52
Gemini-3.8-FlashOpenEvolve156.6430.0085.8545.7147.0872.5654.76
Gemini-3.8-FlashEvoDuet124.3454.9884.4673.6448.1387.8070.33
Gemini-3.8-FlashOpenEvolve854.79100.8184.4062.4871.6487.2282.58
Gemini-3.8-FlashEvoDuet859.6173.4061.5261.8467.6292.8176.82
AuthorsOpenReview ↗

Response by Authors (2/2)

W3 & Q4. Component-wise ablations

A. Thank you for this suggestion. We conducted four component ablations across four tasks using GPT-5.6-Luna. The ablations share a frozen document corpus per task to control retrieval variation.

Component ablationDenoising ↑Sums/Diffs ↑Voyager 2 ↓Swap ↓
Full EvoDuet (paper)0.7019 ± 0.02011.087436 ± 0.0106493.4383 ± 0.00917,894 ± 1,579
w/o knowledge-gap gating0.70971.0864653.44638719,343
w/o query refinement0.70791.0864303.43020617,174
w/o hypothetical evidence scoring0.56871.0734303.44638719,719
w/o persistent search memory0.70001.0864713.44641119,260

In the Table, we observe a drop in performance when a key component is removed.


W4. Predictive utility and document contribution

A. We agree. Figure 8(a) supports a predictive surrogate for query optimization (Spearman ρ = 0.86 across 4,324 retrievals). We additionally conducted a controlled experiment on Denoising and Sums/Diffs with GPT-5.6-Luna. We generated candidate code four times per document combination (ABC, BC, AC, AB). ΔA = mean score(ABC) − mean score(BC), and analogously for B/C.

TaskΔAΔBΔCSpearman ρ
Denoising+0.1554+0.0976+0.0801+0.87
Sums/Diffs0.0000−0.0023−0.0034+1.00

On Denoising and Sums/Diffs, documents with higher predicted scores tended to contribute more to candidate performance. Based on Figure 8(a) and these results, we would like to argue that hypothetical evidence scoring is an informative surrogate for document utility.


Q1. Mean ± SD, actual n, and significance

A. We already report mean ± standard deviation for all 21 tasks in Appendix D.4.

TaskOE: diffED: diffOE: fullED: full
Swap Reduction31.60 ± 44.6915.73 ± 2.9812.16 ± 14.3871.78 ± 8.78
Cassini66.59 ± 7.0597.97 ± 1.4081.10 ± 0.8376.35 ± <0.01
Galileo36.88 ± 0.8937.22 ± 0.3038.26 ± <0.0138.25 ± <0.01
Mariner 1099.99 ± <0.0199.99 ± <0.0199.99 ± <0.0199.99 ± <0.01
Rosetta97.45 ± 0.5661.93 ± 3.4065.27 ± 7.8397.49 ± 3.64
Voyager 298.70 ± 1.8497.40 ± <0.01100.00 ± <0.01100.00 ± <0.01
Denoising0.00 ± 0.0086.05 ± 6.7244.18 ± 62.4762.65 ± 35.42
Domain mixture97.65 ± 0.5997.70 ± 0.5197.42 ± 0.5297.47 ± 0.12
Easy question0.00 ± 0.000.00 ± 0.000.00 ± 0.000.00 ± 0.00
LR/BSZ0.00 ± 0.000.00 ± 0.0047.85 ± 20.5856.59 ± 12.34
Parallel scaling83.80 ± 22.2469.08 ± 1.4398.88 ± 0.9298.16 ± 0.09
AHC0391.89 ± 2.680.00 ± 0.000.79 ± 1.120.00 ± 0.00
AHC05892.17 ± 0.2991.87 ± 1.5192.12 ± 0.6592.44 ± 0.10
Erdős91.66 ± 9.0199.97 ± 0.0199.92 ± 0.0199.92 ± 0.12
AC147.60 ± 7.3940.57 ± 8.5437.39 ± 1.5541.43 ± 29.08
AC240.79 ± 1.1462.16 ± 8.9441.66 ± 42.8869.52 ± 3.52
AC399.64 ± 0.1299.40 ± 0.0599.78 ± 0.0399.59 ± 0.05
CP (n=26)100.00 ± <0.0199.93 ± 0.10100.00 ± <0.01100.00 ± <0.01
CP (n=32)100.00 ± <0.0199.55 ± 0.63100.00 ± <0.01100.00 ± <0.01
Hadamard77.52 ± 18.7575.78 ± 31.6468.58 ± 22.1754.79 ± 0.74
Sums/Diffs16.03 ± 0.0034.15 ± 4.5814.81 ± 1.7230.82 ± 20.92
Backbone / NΔ NDG points95% interval
Qwen 9B, N=1-0.84[-8.12, 6.99]
Qwen 9B, N=8+1.49[-9.38, 11.72]
Luna, N=1+5.56[0.47, 11.86]
Luna, N=8+9.75[1.70, 17.35]
Gemini, N=1+15.57[5.48, 26.16]
Gemini, N=8-5.76[-15.40, 2.15]

At N=1, EvoDuet improves mean NDG over OpenEvolve with both GPT-5.6-Luna and Gemini-3.8-Flash, with exploratory 95% task-bootstrap confidence intervals excluding zero.


Q2. Query refinement with fixed search results

A. We compared Refine and Independent using the same frozen corpus and deterministic BM25 retrieval, so identical queries return identical ranked documents. Refine uses feedback from previous search rounds; Independent generates each query from the episode’s initial context, without feedback. Both used GPT-5.6-Luna, three query rounds, and one candidate per outer iteration.

Refine mean NDG (%)Independent mean NDG (%)Δ (points)
72.5970.86+1.73
Refine achieves a slightly higher mean NDG, although Independent performs better on some tasks.
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer 7rg2,

We understand that you are busy, and we sincerely appreciate the time you have already spent on our paper. As the discussion period is ending soon, we wanted to gently follow up on our response. We would be grateful for any further comments, and we would be happy to address any remaining questions or concerns before the discussion closes.

Thank you again for your time and thoughtful feedback.

Best regards, The Authors

Comments on the paper

6 comments
AuthorsOpenReview ↗

A Brief Note on Supplementary Materials

Thank you for your time and thoughtful feedback. For ease of reference, we would like to share two brief notes:

  • Behavioral trajectory examples (Appendix I): Appendix I presents examples of EvoDuet’s behavioral trajectories. Due to its length, this appendix is provided in the supplementary ZIP file.

  • Experimental statistics (Appendix D): Appendix D reports the mean and standard deviation of the experimental results to help assess their robustness.

We hope these materials are helpful, and we appreciate your consideration.

Dear Reviewers,

The authors have posted their responses. Please take a look and respond to the authors.

Best, AC

Dear Authors,

Just a reminder that you haven’t updated the PDF yet; I can't see any blue marks. Please update it as soon as possible.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.