acceptodds
Under review as a conference paper at ICLR 2027

From Experience to Knowledge: Algorithm Design for Black-Box Optimization with Large Language Models

Abstract

Large language models (LLMs) have shown promise in automated evolutionary algorithm (EA) design through their ability to generate algorithmic code and iteratively refine it based on performance feedback. However, many existing methods rely on iterative search to design algorithms for target problems, requiring repeated LLM calls and candidate evaluations. Moreover, design experience obtained on a target problem may not provide guidance for algorithm design on subsequent problems. To address these limitations, we propose FETK, a three-stage framework that distills problem algorithm design experience into knowledge, enabling an EA architecture to be generated for each subsequent problem with a single LLM call. First, the Experience Collection module collects problem design experience by searching EA architectures on training problems and recording landscape representations, architectural modifications, and performance outcomes. Second, the Knowledge Learning module learns design knowledge by distilling experience across training problems into natural-language guidance and evaluating candidate knowledge updates on validation problems before retaining them. Finally, the Knowledge Reuse module combines the learned knowledge with the landscape representation of each test problem to generate an EA architecture with a single LLM call. Experiments on BBOB and MA-BBOB problems demonstrate competitive performance compared with manually designed EA methods, learning-based design EA methods, and LLM-based design EA methods. https://anonymous.4open.science/r/fetk-DE22

Primary area:
optimization
Keywords:
Automated Algorithm Design, Black-Box Optimization, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BdYdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FETK collects evolutionary algorithm design experience, distills it into natural-language knowledge through validation-gated updates, and generates architectures for new problems using one LLM call. Continuous parameters are subsequently tuned on each target problem. Experiments cover BBOB and MA-BBOB.

Critical strengths
  • Reusing design experience across problems addresses a relevant limitation of iterative LLM-based algorithm design.
  • The evaluation includes component ablations, dimensional transfer, an additional data split, and repeated architecture generation and knowledge learning.
Critical weaknesses
  • End-to-end resource comparability remains unresolved. Appendix D allows 5,000 parameter configurations per test problem, each requiring 1000D1000D objective evaluations. At D=10D=10, this permits 5×1075\times10^7 evaluations before the reported 10410^4-evaluation final run. Baseline-specific design and tuning costs are not fully compared. Reduced deployment-time LLM usage therefore does not establish overall optimization efficiency.
  • The contribution of distilled knowledge is not fully isolated. AO-1 supplies example architectures, but does not test directly selecting a cached training architecture using landscape features and applying identical parameter tuning.
  • Some landscape interpretations lack validation. Ordering samples by nearest-neighbor distance does not ensure spatial adjacency for the reported “autocorrelation”; the additive quadratic coefficient ratio likewise does not generally characterize the full objective’s conditioning.
Questions for authors
  • How do the conclusions change under matched total objective-evaluation budgets, including target-specific tuning?
  • Can landscape-based selection from cached architectures match FETK under identical tuning, and does knowledge validation itself include parameter tuning?
Reviewer_FShCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FETK which learns reusable design knowledge from EA design experience and reuses it to generate EA architectures for black-box optimization problems. The authors introduce an Experience Collection module to collects problem design experience, a Knowledge Learning module to distill collected design knowledge into natural-language guidance, and a Knowledge Reuse module to generate an EA architecture within single LLM call. Experimental results show that FETK generates effective EAs.

Critical strengths
  1. FETK uses both success-driven and failure-driven experiences. This improves data utilization and make the training more comprehensive.

  2. In the knowledge reuse stage, FETK generates a EA architecture with a single LLM call. The token consumption comparison shows that FETK-1 uses substantially fewer tokens than iterative NK-10 while achieving comparable performance.

Critical weaknesses
  1. The testing instance set is limited. A small set (6 BBOB + 8 MA-BBOB in total) may not comprehensively evaluate the proposed framework. The principle of selecting training, validation and testing instance set is not reported. Besides, the tested instances are all synthetic problems, they are in the same distribution. Validating FETK's performance on out-of-distribution problems such as some real-world application problems could enhance the generalization analysis on different landscape characteristics.

  2. The included baselines are biased. Other LLM-based EA generation methods such as EoH[1] and LLaMoCo[2] is not compared. LLaMoCo is similar to FETK who also generates EA codes in one single LLM call. State-of-the-art manually designed EAs such as CMA-ES, DE and PSO, non-LLM-based automatic EA design methods such as GSF[3] and ALDes[4] are not included.

  3. The GA based parameter searching method in the Knowledge Reuse module may have large computational overhead.

[1] Fei Liu, et al. "Evolution of heuristics: towards efficient automatic algorithm design using large language model". In Proceedings of the 41st International Conference on Machine Learning (ICML'24), Vol. 235. 32201–32223, 2024.

[2] Ma, Zeyuan, et al. "Llamoco: Instruction tuning of large language models for optimization code generation." IEEE Transactions on Evolutionary Computation (2026).

[3] Yi, Wenjie, et al. "Automated design of metaheuristics using reinforcement learning within a novel general search framework." IEEE Transactions on Evolutionary Computation 27.4 (2022): 1072-1084.

[4] Zhao, Qi, et al. "Automated metaheuristic algorithm design with autoregressive learning." IEEE Transactions on Evolutionary Computation (2024).

Questions for authors
  1. What does the rejected-update memory contain? How to update the memory?

  2. Although the paper reports token consumption, could the authors provide a wall-clock time overhead comparison?

Reviewer_it57OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The key novelty is the one-shot algorithm design of an evolutionary algorithm by using LLMs and knowledge learning. A smart way to extract knowledge from algorithms is introduced and used to support the LLM in one-shot algorithm design. The approach is tested on BBOB and MA-BBOB, with good results, seemingly outperforming existing loop-based approaches.

Critical strengths
  • A novel approach towards extracting knowledge from the generated algorithms, then using this knowledge to enhance the LLM's capability to generate optimization algorithms in one shot.
Critical weaknesses
  • The presentation is really bad, therefore the paper is difficult to follow. The math notation does not make it easier, as many notational aspects are only vaguely introduced and do not help understanding.
  • Most recent competitive approaches using an evolutionary loop are not considered in the experimental comparison. For example, LLaMEA wirth population sizes larger than one.
Questions for authors

I find Fig. 1 hard to understand. What are you trying to say? What is a problem here, do you mean instances of problem classes? What is cross-problem reuse then? This is very unclear. The reuse across different problem types seems to imply the search for the general-purpose solver that can solve any kind of problem - this is exactly what does not work in reality. Are you proposing that the LLM learns across different problem classes and then will be able to generate the optimal algorithm, given a problem instances from a specific class, in one shot, without iterations? Clarify! 092: What is "architecture search across training problems"? Which architecture space are you talking about? 143, 144: What are "directed connections between operators", and "information-routing ratios"? What are the domains of E and R? 147: What is E_train? Eqn. (1): What is "Agg"? Fig 2: What does an architecture generation LLM do? Generate an algorithm's "architecture"? What is that? 185: Ok, here is becomes a bit clearer you mean EA architectures. 215: What do you mean here by "ELA calculation"? If it is really ELA features, which seems to be the case, then clarify that here, and why we need them! What is K_e_j? Please define K more clearly. 303: Using a GA to optimize hyperparameters is not state of the art. You should use a hyperparameter Optimizer such as SMAC3, for example. Overall, the math formalism is Only partially helpful, as it is clumsy, not easy to understand, and often Only vaguely defined. This Needs to be significantly improved. Comparison to more state of the art Evolutionary LLM based approaches should be done, too. There are now population-based versions of LLaMEA, for example.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.