acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Exploratory Landscape Analysis for LLM-based Meta-Black-Box Optimization

Abstract

Large Language Models (LLMs) have recently advanced Meta-Black-Box Optimization (MetaBBO) by enabling optimizer design with reduced expert intervention, while existing approaches typically rely on fixed optimization-state representations, such as predefined Exploratory Landscape Analysis (ELA) indicators. We argue that adaptive optimization-state perception is more effective than fixed optimization-state representations for improving MetaBBO performance. Motivated by this idea, we propose AdaELA, a plug-and-play adaptive ELA selection framework for LLM-based MetaBBO. AdaELA dynamically selects informative ELA indicators through a population-aware self-attention encoder and a cross-attention-based selection mechanism, and provides the selected landscape information to LLM-based meta-policies for optimizer control. We formulate the training of AdaELA as a multi-task optimization problem and optimize its parameters using a neuroevolution strategy. Extensive experiments demonstrate that AdaELA consistently improves different LLM-based MetaBBO methods, generalizes to unseen optimization tasks, and enables further performance improvements through fine-tuning. This work highlights adaptive optimization-state perception as a promising direction toward more generalizable and effective MetaBBO.

Primary area:
optimization
Keywords:
Meta-Black-Box Optimization, Large Language Model, Exploratory Landscape Analysis, Adaptive Optimization-State Perception

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yGLhOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2(+1)
Presentation
2(+1)
Soundness
3(+2)
Read review
Paper summary

This paper looks at exploratory landscape analysis (ELA) for black box optimization when using LLM-driven meta-black-box optimizers. It proposes AdaELA, a method to adaptively select ELA indicators for use by LLM meta-optimizers. It claims to see the learned settings for AdaELA generalize to out of distribution optimization tasks.

The paper suffers from lack of readability, general vague terminology and lack of concrete descriptions for its methodology.

Critical strengths

The paper does substantive analysis of exactly what role the ELA plays in in the ELA + LLM metaBBO, extracting exactly what benefit it provides.

Critical weaknesses

The paper generally reads as unclear, and is generally difficult to understand what exactly is being tested and how. I find it vague, dense, and confusing to read.

I find it difficult to ascertain how LLMs are being used in the BBO problems. There is no indication of what a prompt to this LLM would look like, how it would respond, what the harness looks like, what tools it may or may not have access too. The only mention of a prompt is in the appendix and it is vague. Additionally, given this is black box optimization, it is missing one comparison I think would be necessary for this type of analysis, which is simply providing codex or claude code with the optimization problem and measuring the effectiveness with the /goal command.

Revisions

See below comment for my revised interpretation of the paper.

Questions for authors

What is the role of the LLM in the BBO setups? What does the harness look like? What tool calls or methods does the LLM have for interacting with this system? In figure 1, where does the LLM fit in? What do LLM rollouts look like? What is the information passed to the MetaBBO?

3 comments

We sincerely apologize that our work has caused confusion. We have responded to your questions in detail and provided further explanations of our method. We sincerely hope you will reconsider our work in light of our response.


Clarification: What AdaELA Does

We would first like to clarify the role of AdaELA.

  1. AdaELA is a plug-and-play ELA selector, not an LLM-based MetaBBO method. It is a lightweight attention-based network trained to select ELA indicators (as stated in Lines 90–101 of our manuscript); it is not an agent and does not itself include an LLM, a harness, or tool calls.
  2. Existing methods often use a fixed set of ELA indicators throughout optimization. AdaELA instead selects indicators according to the current optimization state, providing the LLM with information suited to different search stages.
  3. Our focus is on whether adding AdaELA improves existing MetaBBO methods. The underlying LLMs and their execution workflows (e.g., prompts, harnesses, and tool use) belong to those methods.

How LLMs Are Used in MetaBBO

Different optimization problems and stages may require different search strategies. MetaBBO uses a meta-policy to observe the search process and guide the optimizer accordingly[1,2,3]. In LLM-based MetaBBO, the LLM serves as this meta-policy. To make the process easier to understand, we provide an illustration in https://anonymous.4open.science/r/ICLR2027-EBB5/LLM-based%20MetaBBO.pdf.

  1. Prompt construction. As stated in Section 3.4, AdaELA is a plug-and-play module that can be integrated into different LLM-based MetaBBO methods without changing their optimization frameworks. It provides the LLM with selected ELA indicators, their values, and semantic descriptions as optimization-state information. To clarify this process, we provide an illustration in https://anonymous.4open.science/r/ICLR2027-EBB5/AdaELA%20prompt%20.pdf.
  2. What an LLM Rollout Looks Like.
    • A typical rollout in LLM-based MetaBBO is one complete optimization run. At each decision step, the LLM receives a prompt containing the current optimization information and returns a decision to guide the optimizer. The optimizer then evaluates new solutions and updates its state. This process repeats until the evaluation budget is exhausted.
    • AdaELA follows this general workflow. When it is used, the selected ELA indicators, their values, and descriptions are included in the prompt at each decision step.
    • If you feel it would be helpful, we will expand the introduction to these basic concepts to make our work easier to understand for readers unfamiliar with MetaBBO.

Comparison with Codex or Claude Code

  1. AdaELA is a plug-and-play module, not a complete optimizer. Our goal is to evaluate whether adding AdaELA improves existing LLM-based MetaBBO methods by adaptively selecting the optimization information provided to the LLM.
  2. Codex and Claude Code, on the other hand, are general-purpose coding agents that can use their own tools and workflows to solve a given task. Therefore, directly comparing AdaELA with Codex or Claude Code would not be appropriate.

Summary

  1. AdaELA is a plug-and-play state perception module for LLM-based MetaBBO. It is a lightweight attention-based network that selects ELA indicators; it is not an LLM or a complete optimizer.
  2. Our central finding is that adaptive optimization-state perception can improve MetaBBO performance. Motivated by this idea, we propose AdaELA to provide LLM-based methods with state information that changes with the optimization process.
  3. AdaELA supports existing methods. It adapts the ELA information provided to the LLM and can be integrated into different MetaBBO methods. Our experiments compare these methods with and without AdaELA to assess its contribution.
  4. The learned selector is transferable. Once trained, AdaELA can be applied to unseen optimization problems and MetaBBO methods without additional training.
  5. To help you better understand the effect of AdaELA, we use CMA-ES as a baseline and compare each MetaBBO method with and without AdaELA. We report the mean objective value of CMA-ES minus that of the compared method; a larger value indicates better performance. The results are available at https://anonymous.4open.science/r/ICLR2027year-0C5F/result.pdf.

We apologize again for the lack of clarity and sincerely hope you will reconsider our paper in light of our response. We wish you all the best!


  • [1] Ma Z, et al. Toward automated algorithm design: A survey and practical guide to meta-black-box-optimization[J]. TEVC, 2025
  • [2] Ma Z, et al. Metabox: A benchmark platform for meta-black-box optimization with reinforcement learning[C]. Advances in Neural Information Processing Systems, 2023
  • [3] Yang X, et al. Meta-Black-Box optimization for evolutionary algorithms: Review and perspective[J]. Swarm and Evolutionary Computation

I would like to apologize to the authors on an earlier reading placing too much focus on the lack of details of the BBO methodology, given the focus of this paper on the ELA. Not understanding how the LMM-based MetaBBO worked and not seeing an explanation within the paper was functioning led to confusion in my reading of the paper - but I see there is already existing literature on these. I have followed through the cited LLM-based MetaBBO papers, and while this paper does not need to have an extensive description of their setup, I do believe it still requires more description than is given now, in the appendix - and would like to see example LLM interactions (prompt/ responses). I also think possibly a figure and/or prompt/response in the related works section of how the LLM-based MetaBBO works could be helpful, though my score does not depend on this.

I was initially concerned that results would struggle to generalize to MetaBBO's with newer models. For example, autoEP runs with Qwen-3-30B and gpt-4, these models are well behind the frontier. Newer models tend to be able to take advantage of more information - I had not caught the first time that late in the appendix there is a chart covering this with newer models. This was an immediate concern on reading the paper, and either moving that section up, or providing a small table with commentary on this in the main paper would be helpful.

I am revising my score to a 3.

We sincerely thank you for your recognition of our work and for taking the time to consider our responses.

We wish you all the best!

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

This paper proposes AdaELA, an adaptive Exploratory Landscape Analysis (ELA) indicator selection module for LLM-based Meta-Black-Box Optimization (MetaBBO). Instead of using a fixed optimization-state representation in the prompt, AdaELA uses a lightweight attention network (~5k parameters) meta-trained via FCMAES to dynamically select a subset of 25 candidate landscape indicators along with their natural-language descriptions at each iteration. Experiments across multiple LLM-based MetaBBO baselines on BBOB, BBOB-NOISY, and an aircraft design problem, together with ablation studies, are conducted to evaluate its effectiveness and transferability.

Critical strengths

This paper addresses optimization-state representation in LLM-based MetaBBO and introduces a lightweight, plug-and-play indicator selection module with negligible inference overhead. The experimental evaluation is well structured, testing transferability across both unseen optimization problems and unseen MetaBBO algorithms with consistent relative improvements.

Critical weaknesses
  1. The methodological novelty is relatively incremental, as it mainly combines a standard attention-based selector with existing LLM-based MetaBBO frameworks, placing the primary burden of contribution on empirical validation. However, the experiments do not adequately isolate the effect of dynamic selection from the newly introduced 25-indicator pool, the natural-language semantic descriptions, and step-to-step prompt variation. In Figure 3, Random ELA and All ELA already achieve substantial gains over the original baselines, and there is no comparison against the best static indicator subset optimized under a comparable search budget.

  2. The evaluation protocol and metric reporting make it difficult to assess the practical magnitude and generality of the improvements. Performance is reported solely as relative exp⁡(mean Z-score)\exp(\text{mean Z-score}) normalized by each baseline's standard deviation, without absolute objective values, statistical significance tests, or standard non-LLM optimizers (such as CMA-ES) as a reference point. Additionally, the exact train/test split of BBOB functions and the total meta-training cost in terms of LLM calls and runtime are not clearly documented, which weakens the claims of generalization and practical efficiency.

Questions for authors

Q1. Compared with the original methods, AdaELA introduces a 25-indicator library, semantic descriptions, and dynamic selection simultaneously. How does AdaELA perform against the best fixed subset of the same indicator library (with identical semantic descriptions) found under a comparable search budget, and how much of the improvement comes from the semantic descriptions or prompt variation alone?

Q2. Since Z-score normalization can amplify small absolute differences when baseline variance is low, could the authors report absolute objective values, statistical significance tests, and a standard optimizer such as CMA-ES under the same 4,000-evaluation budget to clarify the practical effect size?

Q3. Which specific BBOB functions and instances were used for meta-training versus held-out testing, and what is the total computational and LLM-call cost required during the FCMAES meta-training phase?

2 comments

Q1: Effects of Dynamic ELA Selection, Semantic Descriptions, and Prompt Variation

1. Comparison with an Optimized Fixed ELA Subset

We appreciate this suggestion.

  • We had considered this issue before. However, the 25-indicator library contains 2252^{25} possible subsets, making it difficult to efficiently identify the globally optimal combination. Moreover, a subset that works best for one MetaBBO method may not be optimal for others, making it difficult to find a single fixed subset that works best across methods.
  • To address your concern, we follow the feature selection idea in [1] and use AdaELA’s meta-training objective (Eq. 6) to search for a fixed ELA subset on BBOB across AutoEP, OPRO, and LAOS, using the same number of evaluations as AdaELA’s meta-training. We then test this subset on BBOB-NOISY. As shown in https://anonymous.4open.science/r/ICLR-2027-year-73C1/AdaELA_vs_Optimized_Fixed_BBOB_NOISY.pdf, its relative performance is below 1 on all tested methods, using AdaELA as the baseline. This supports the benefit of adaptive ELA selection.

2. Effect of Semantic Descriptions and Prompt Variation

We understand the reviewer’s concern. The performance gain may involve three factors:

\leftarrow\text{Dynamic ELA selection}+ \text{Semantic descriptions}+\text{Prompt variation}$$ - However, we think their effects are not simply additive. These factors are closely coupled: changing the selected indicators also changes their descriptions and the prompt provided to the LLM. Their interactions make it difficult to precisely separate the contribution of each factor. - For **prompt variation**. AdaELA works by adaptively changing the ELA indicators in the prompt. Our `Random ELA` experiment also changes the prompt, but performs worse than AdaELA. This suggests that the improvement comes from how the indicators are selected, rather than prompt changes alone. - To study the role of **semantic descriptions**, we conducted two experiments: - We re-trained AdaELA using FCMAES, providing only ELA indicator names and values without semantic descriptions. AdaELA converged quickly, but the selected indicators remained largely unchanged during optimization, and there was no clear performance gain. One possible explanation is that FCMAES favors indicators whose names are easier for the LLM to understand, such as `fitness_std` and `population_diversity`. As shown in **https://anonymous.4open.science/r/ICLR-2027-year-73C1/Without%20semantic%20descriptions.pdf**. - We also tested the default AdaELA with its semantic descriptions **removed or replaced**. Removing them led to worse performance, while replacing them had little effect. As shown in **https://anonymous.4open.science/r/ICLR-2027-year-73C1/AdaELA_semantic_descriptions.pdf** - These results suggest that semantic descriptions are important, while AdaELA is robust to changes in the wording of semantic descriptions. ----- ## Q2: Practical Performance Improvements of AdaELA 1. We agree that a small baseline variance may amplify differences in the Z-score. To address this concern and provide a comparison with a standard optimizer, we use CMA-ES as a reference and report the following difference on each BBOB problem: $$\Delta f =\text{Mean objective value of CMA-ES}-\text{Mean objective value of the compared method}.$$ We report this difference for each MetaBBO method, both with and without AdaELA. For these minimization problems, a **positive value** indicates that the compared method achieves a lower mean objective value than CMA-ES. The detailed results are available in **https://anonymous.4open.science/r/ICLR2027year-0C5F/result.pdf**. 2. Across 24 problems and six MetaBBO methods, AdaELA achieves statistically significant improvements in 133 of 144 cases, according to the Wilcoxon signed-rank test. ----- ## Q3: Meta-Training, Testing and Computational Cost 1. **Meta-Training and Testing Setup**. - As stated in the **Experimental Setup in Section 4.1**, we use all BBOB instances for meta-training and evaluate transfer on BBOB-NOISY and HPA. The **meta-training methods** are AutoEP, OPRO, and LAOS, while EvoLLM, LEO, and PAIR are **held out for testing**. - This allows us to examine transfer across both **seen and unseen problems and algorithms**. 2. **Training Cost**. - During default meta-training, BBOB suite is evaluated 45,000 times across the three MetaBBO methods. - In our previously experiments, we also evaluated a lower-budget setting with *Q=5* and 50 FCMAES iterations. As shown in **https://anonymous.4open.science/r/ICLR-27-2C2C/AdaELA_budget_BBOB_NOISY.pdf**, this setting uses approximately **16.7%** of the default training compute while retaining **85–95%** of its performance. ----- **We hope our responses address your concerns, and we wish you all the best!** ----- - [1] Mersmann O, et al. Exploratory landscape analysis[C]. GECCO, 2011.

Thank you again for your time and helpful feedback. As the discussion deadline is approaching, we would appreciate it if you could review our responses and additional results when you have a chance. Please let us know if any concerns remain or if further clarification would be helpful.

Thank you for your consideration.

Reviewer_paziOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a plug-and-play adaptive ELA selection framework for LLM-based MetaBBO. It first employs a population-aware self-attention encoder and a cross-attention mechanism to dynamically select informative ELA indicators during optimization. The selected ELA indicators and descriptions are then prompted to the LLMs to control the downstream optimization. The proposed framework is formulated as a multi-task optimization problem and optimized via a neuroevolution strategy. Evaluation is conducted on the BBOB test suites in the COCO platform.

Critical strengths
  1. The idea of introducing ELA to metaBBO is interesting.
  2. The zero-shot generalization performance is good.
  3. The plug-and-play design enhances the applicability of the proposed framework.
Critical weaknesses
  1. Evaluating a single AdaELA parameter configuration requires the entire LLM to perform numerous independent runs across the multi-task space. Even with parallelization, the cost of this neuroevolution-based training remains high.
  2. The ablation studies indicate that if model complexity is increased (e.g., by adding attention heads or hidden layer dimensions), AdaELA will overfit to improve performance on specific tasks, and its cross-task generalization ability will drop. Could the authors provide some in-depth understanding of this tradeoff?
  3. Although the neural network inference is fast, in the current implementation, the computation of the massive ELA indicator library is executed entirely serially, which constitutes the largest source of time overhead for the framework in actual deployment. Could the authors provide a detailed computational analysis of each component?
  4. The input state (ELA combination) changes dynamically, which seems to be hard to apply to the RL-based metaBBO methods.
Questions for authors

Please refer to the weakness above.

2 comments

W1: Meta-training cost

We agree that meta-training AdaELA is computationally expensive. However:

  1. It is primarily a one-time cost: the trained model can be transferred to unseen problems and MetaBBO methods without retraining.
  2. In our previously conducted experiments, we also evaluated a lower-budget configuration with Q=5Q=5 and 50 FCMAES iterations, compared with the default setting of Q=15Q=15 and 100 iterations. As shown in https://anonymous.4open.science/r/ICLR-27-2C2C/AdaELA_budget_BBOB_NOISY.pdf, the lower-budget configuration uses approximately 16.7% of the default training compute while retaining 85–95% of its performance.
  3. High computational cost is a common challenge for learning-based ELA methods [1–2]. Despite this limitation, our main finding is that adaptively selecting state information can improve MetaBBO performance.

W2: Model complexity and generalization

We acknowledge that this trade-off was not sufficiently explained.

  1. In RQ6, “unseen tasks” refers to optimization problems excluded from meta-training. A more complex model may capture task-specific patterns that transfer poorly to these problems.
  2. On the Human-Powered Aircraft Design (HPA) problem, we compare AdaELA under three meta-training configurations: BBOB alone, BBOB with BBOB-NOISY, and BBOB with both BBOB-NOISY and HPA. For each configuration, we evaluate two architectures (h=16h=16 and h=64h=64, both with three attention heads), using the default AdaELA architecture (h=16h=16, one head) as the reference baseline. As shown in https://anonymous.4open.science/r/ICLR-27-1-7B72/AdaELA_model_complexity_HPA.pdf, adding BBOB-NOISY improves both models’ performance on unseen HPA, suggesting enhanced transferability. When HPA is included in meta-training, it becomes a seen problem, and the more complex model performs better, consistent with our RQ6 finding that greater model complexity can benefit performance on specific tasks.
  3. A more complex model can provide modest performance gains but may increase meta-training cost. This observation is consistent with previous studies [1,3]. Considering this trade-off, we choose the smaller model as our default.

W3: ELA computational overhead

We agree that computing ELA indicators is the main source of runtime overhead.

  1. Appendix B.4, Table 2 reports the runtime of each AdaELA component. For example, with N=200N=200 and D=15D=15, ELA computation, self-attention, cross-attention, and the full pipeline take 413.536 ms, 3.344 ms, 0.132 ms, and 420.115 ms, respectively.

  2. As discussed in the appendix, computing ELA indicators in parallel may reduce runtime. We tested a parallel version using five processes and achieved a 1.56–1.82× speedup in the settings below. A better parallel implementation may further improve speed, although this is not the main focus of our work.

NNDDSerial (ms)Parallel (ms)Speedup
50528.4117.671.61×
5015272.47161.541.69×
100533.3918.311.82×
10015317.11189.771.67×
200552.7033.871.56×
20015413.31254.301.63×

W4: Applicability to RL-based MetaBBO

We agree that dynamically changing ELA combinations are challenging for RL-based methods. This is one reason why we focus on LLM-based MetaBBO, although we also conduct proof-of-concept experiments with RL-based methods.

  1. Why LLM-based MetaBBO? Conventional RL policies typically require fixed-dimensional inputs, making it difficult to directly use changing ELA combinations. In contrast, LLMs can accept different combinations as text without retraining.

  2. RL-based experiments. In Appendix C, we evaluate adaptive ELA selection with LDE, DEDDQN, and RL-DAS. For each encountered ELA combination, we train a corresponding RL policy. The results suggest that adaptive ELA selection can also benefit RL-based MetaBBO, although training multiple policies is computationally expensive.

  3. Future direction. As discussed in the appendix, Hierarchical Reinforcement Learning may provide a more efficient way to handle changing ELA combinations. Our team is currently exploring this direction.


We thank the reviewer for the constructive comments. We hope our responses address your concerns, and we wish you all the best!


  • [1] Ma Z, Chen J, Guo H, et al. Neural exploratory landscape analysis for meta-black-box optimization[C]. ICLR, 2025.

  • [2] Seiler M V, Kerschke P, Trautmann H. Deep-ELA: Deep exploratory landscape analysis with self-supervised pretrained transformers for single- and multiobjective continuous optimization problems[J]. Evolutionary Computation, 2025.

  • [3] Du Y, Yu H, Jiang J, et al. Meta-black-box optimization can do search guidance for expensive constrained multi-objective optimization[C]. ICML, 2026.

Thank you again for your time and helpful feedback. As the discussion deadline is approaching, we would appreciate it if you could review our responses and additional results when you have a chance. Please let us know if any concerns remain or if further clarification would be helpful.

Thank you for your consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Our Responses

We thank the reviewers for their comments. In our responses, we have:

  1. Clarified our contribution: AdaELA is a plug-and-play module that adaptively selects optimization-state information for existing LLM-based MetaBBO methods. (Reviewer yGLh)

  2. Examined adaptive selection: We compared AdaELA with an optimized fixed ELA subset and studied the effects of semantic descriptions and prompt variation. (Reviewer rgHh:Q1)

  3. Expanded the evaluation: We provided objective values, statistical tests, and comparisons with CMA-ES to make the performance improvements easier to assess. (Reviewer rgHh:Q2) (Reviewer yGLh)

  4. Analyzed computational cost: We reported the meta-training cost, evaluated a lower-budget configuration, and tested parallel ELA computation. (Reviewer pazi: W1, W3) (Reviewer rgHh:Q3)

  5. Clarified generalization and application: We explained the training and testing setup, examined the effect of model complexity and training diversity, and discussed how AdaELA relates to RL-based MetaBBO. (Reviewer pazi: W2, W4)

  6. Explained the workflow: We provided additional illustrations and explanations of how AdaELA integrates with existing LLM-based MetaBBO methods. (Reviewer yGLh)


The additional results and illustrations provided in our responses are summarized below.

ItemPurposeLink
Lower-budget meta-trainingEvaluate AdaELA under a reduced meta-training budgetResults
Model complexity and training problemsExamine the effects of model complexity and training problem diversityResults
Optimized fixed ELA subsetCompare AdaELA with an optimized fixed subset of ELA indicatorsResults
Training without semantic descriptionsExamine AdaELA’s selection behavior when trained without semantic descriptionsResults
Removing or replacing semantic descriptionsEvaluate the effect of semantic descriptions and alternative wordingResults
Objective values and CMA-ES comparisonProvide objective-value results using CMA-ES as a referenceResults
LLM-based MetaBBO workflowIllustrate the role of the LLM as a meta-policy in MetaBBOIllustration
AdaELA prompt constructionIllustrate how selected ELA information is included in existing MetaBBO promptsIllustration

We hope these responses address the reviewers’ concerns and clarify our contribution. We welcome any further questions or suggestions. After the rebuttal period, we will consolidate all feedback and revise the manuscript accordingly.

The reviewer has provided a point-by-point response to your comments; please respond at your earliest convenience.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.