acceptodds
Under review as a conference paper at ICLR 2027

Principled Data Generation for MetaBBO via Active Task Selection

Abstract

Meta-Black-Box Optimization (MetaBBO) aims to acquire generalized optimization strategies by training on extensive sets of functions. While traditional approaches predominantly rely on fixed benchmarks, recent efforts have shifted towards maximizing landscape diversity to broaden training distribution coverage. However, we identify a critical Diversity-Quality Gap: Simply increasing landscape variety often yields tasks that are misaligned with the agent's evolving capabilities. To address this, we propose Hierarchical Active Task Selection (HATS), a principled framework that constructs an automated curriculum based on the agent's potential for improvement. HATS introduces a regret-based utility metric measuring the performance gap between the current policy and a competent baseline to guide a bi-level selection process. Specifically, it combines a multi-armed bandit for dynamic function class weighting with an active sampler for parameter replay. Experiments demonstrate that HATS significantly outperforms diversity-driven baselines and achieves superior generalization on real-world tasks with fewer training steps, highlighting the value of quality-based data generation in MetaBBO.

Primary area:
optimization
Keywords:
MetaBBO, Data Generation, Curriculum Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rLAYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HATS selects training tasks for meta black-box optimization using the performance gap between the learner and a baseline optimizer. A bandit selects function classes, while prioritized replay selects instances. Experiments cover multiple optimization backbones and synthetic and real-world tasks.

Critical strengths
  • The method addresses a useful gap between task diversity and training value, offering a practical curriculum for learned optimizers.
  • The method improves results across multiple optimization backbones and task types.
Critical weaknesses
  • The utility proxy is not directly validated against training value. The diversity-quality study does not show that teacher-relative gaps predict learning progress, leaving the central selection principle insufficiently supported.
  • The efficiency evidence is limited to training episodes. It is unclear how task-screening trajectories and reference-optimizer evaluations are counted, so the benefit under a matched total evaluation budget remains uncertain.
Questions for authors

See the weaknesses above.

3 comments

Thanks for your encouraging comments. Below please find our responses.

W1 Validation of the utility proxy

Thanks for your helpful comments. We agree that the diversity–quality study alone does not show that teacher-relative gaps predict learning progress. To validate whether this proxy translates into actual learning gains, we conduct two illustrative experiments accordingly.

Firstly, we visualize the relationship between different scoring metrics and LOO quality in https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/loo_validation_3panel.pdf. We start from a fixed mid-training checkpoint, randomly generate 32 instances, and compare three metrics: 1) our proposed teacher-student gap utility in Eq. (4), 2) the landscape latent diversity in Eq. 2, and 3) random score. Here we can observe that our proposed utility exhibits a stronger correlation with LOO quality (Pearson r=0.734r=0.734, Spearman ρ=0.675\rho=0.675) than latent diversity (r=0.184r=0.184, ρ=0.222\rho=0.222). Its Spearman correlation also exceeds the central 95% range of correlations obtained from 200,000 random scores ([−0.350,0.349][-0.350, 0.349]), whereas the diversity correlation falls within this range. This provides clear evidence that our proposed teacher-student performance gap utility offers a better predictive signal towards the instances' quality than the diversity-based scoring.

Additionally, we directly examine whether the utility predicts subsequent learning gains through controlled continuation experiments. Specifically, we resume from checkpoints at epochs 25, 50, and 75 and group tasks according to their teacher–student log-gap utility UU : too-easy (U<−0.5U<-0.5), in-zone ( −0.5≤U≤4.0-0.5\leq U\leq4.0 ), and too-hard ( U>4.0U>4.0 ). For each checkpoint, we randomly select 10 tasks from each zone and continually train each model for 1,000 episodes. The following table shows the performance gains over the base policy on held-out test sets, where the in-zone tasks yield the largest gains at all three checkpoints.

Start ckptEasyIn-zoneHard
Epoch 255.05%26.00%15.33%
Epoch 5016.51%18.26%−5.31%
Epoch 751.31%4.82%−7.29%

These results provide supportive evidence that the teacher-relative gaps offer predictive signals towards better instance quality. We have added these experiments to Section 5.2 (RQ5, Figure 5) and Appendix D.1 (Table 3). Thank you very much.

W2 Evaluation under matched budgets

Thanks for your good question. It is worth comparing different methods under the same training-time fitness evaluation budgets, which include both learner and teacher-baseline evaluations. The additional evaluation budgets for different MetaBBO backbones are different:

  • For SYMBOL [1], there is no extra budget beyond the original method as SYMBOL typically uses a teacher method to guide the student policy, naturally producing teacher evaluation. In HATS implementation, we directly reuse the teacher trajectories already required by SYMBOL's training objective to compute utility, without additional fitness evaluation budgets.
  • For other teacher-free backbones like LDE [2] and B2Opt [3], the teacher baseline indeed introduces more budgets. Hence, to make a fair comparison, we train DR and Static to their final checkpoints, compute their total training FEs, and select the corresponding HATS checkpoint saved at the matched total training FE budget (summing both learner and teacher evaluations) for testing on unseen CEC instances. As shown in https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/matched_budget_comparison.pdf, the results are consistent that HATS > DR > Static. These results support the benefit of HATS even under the same evaluations on training instances.

We have added these experiments and checkpoint-alignment details to Section 5.2 (RQ2) and Appendix D.7 (Figure 11). Thank you very much.

[1] Symbol: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning. ICLR 2024.

[2] Learning adaptive differential evolution algorithm from optimization experiences by policy gradient. IEEE TEvC 2021.

[3] B2opt: Learning to optimize black-box optimization with little budget. AAAI 2025.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

Thanks for the detailed responses and additional experiments. The new results address my concern. I am keeping my weak acceptance recommendation.

Dear Reviewer rLAY,

Thanks for your feedback! We are glad to hear that your concerns have been addressed. We sincerely appreciate the time and effort you have dedicated to reviewing our paper and providing thoughtful and valuable comments.

Best regards,

Authors of Submission 712

Reviewer_M8sNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose Hierarchical Active Task Selection to address the identified Diversity-Quality Gap in MetaBBO. The authors use a multi-armed bandit to allocate training budget across optimization problem classes and a utility-aware replay method to sample instance parameters. The framework is evaluated with multiple MetaBBO backbones on CEC, BBOB, Protein Docking and UAV benchmarks. The reported results show that the proposed framework improves generalization and training efficiency of MetaBBO methods.

Critical strengths
  1. The observation of the Diversity-Quality Gap in MetaBBO reveals that simply expanding training problem set with randomly generated instances with different features may not be benefit to training efficiency.

  2. The experimental results show that the proposed framework improve the generalization performance of SYMBOL on CEC, BBOB, Protein-Docking and UAV benchmarks. The evaluation results on LDE, B2OPT and GLHF also validate the effectiveness of the proposed method.

Critical weaknesses
  1. The measurement of the Regret Gap depends on the chosen baseline optimizer. Since optimizers like MadDE may not be optimal for all instances, it may bias the learned policy. HATS ensures the agent learns only from tasks with high utility, this could bias the learned policy toward imitating with MadDE behavior, rather than developing broader optimization capabilities on instances where the agent already surpasses the teacher.

  2. Experimental settings are not detailed, such as the dimensions of the problems, the number of the task classes, and how to sample parameters from generator.

  3. Layout: Most of the equations in the paper are not numbered. Figure 1(b) is first mentioned in Section 4, which is four pages away from the figure. Place the figure and the corresponding main paper text together could improve the readability.

Questions for authors
  1. How sensitive is HATS to the selection of the baseline optimizer in Eq.(1)? What is the difference to use PSO or CMA-ES?

  2. What is the motivation of adding a small constant to the softmax in Section 4.1? how to determine the value of the constant? It seems that the number of classes (K) could not exceed 20, otherwise the probabilities will be reversed?

  3. What is the number of classes? How to divide the problem instanses into classes? Is the division the same across benchmarks such as CEC and BBOB?

  4. What is the paremeter generator in Algorithm 1, line 5?

5 comments
AuthorsOpenReview ↗

Response by the authors

Thanks for your helpful comments. Below please find our responses.

W1 & Q1 Sensitivity to the baseline teacher optimizer

Thanks for the valuable comment. Following your suggestion, we apply multiple teachers (CMA-ES, PSO, and MadDE) to multiple MetaBBO backbones. As shown in https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/teacher_sensitivity_rank_bar.pdf, HATS is reasonably robust across different teacher choices, while the overall trend of teachers is consistent across backbones: using MadDE as the teacher generally obtains the best average rank.

A possible reason is that the teacher should also be in the learnable zone for the current agent. A much stronger teacher (e.g., CMA-ES) may make too many tasks look overly hard, while a much weaker teacher may make the utility less informative. From the results, MadDE is slightly weaker than CMA-ES/PSO, but it remains closer to the agent’s capability, yielding a better curriculum.

We also examine your concern in W1 about whether teacher-based selection restricts further learning once the agent surpasses the teacher. We replace the baseline optimizer with random search at the epoch-5 checkpoint and continue training for 20 epochs, as shown in https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/teacher_random_3panel.png. Although the batch-average log-gap becomes negative, the agent’s performance continues to improve, and MAB entropy does not collapse. These results show that the agent does not imitate the teacher's behaviors; instead, the agent can continue improving after surpassing the sampling reference. In our HATS framework, the teacher is only used to prioritize instances that the agent has not mastered, rather than being distilled. Thus, the behaviors of the student agent would not stick to the teacher.

Overall, these results highlight that teacher selection is important while confirming that HATS continues improving even after surpassing a weak reference. We thank the reviewer for pointing this out; we have added these experiments and clarified this discussion in Appendix D.2 (Table 4 and Figure 6), and also noted adaptive teacher selection as future work in Section 6. Thank you very much.

W2 & Q3 & Q4 Not detailed experimental settings

Thank you for pointing this out. For training, our CEC set covers 10 classes of 10D functions, while the BBOB set includes 24 classes of 10D functions. These two settings follow recipes from SYMBOL [1] and MetaBox [2], respectively.

For evaluation on CEC / BBOB, we use held-out instances with fixed seeds for reproducibility. The standard synthetic test setting also uses dimension d=10. For real-world tasks such as Protein and UAV, we use MetaBox split difficulty=all and their problem-specific dimensions (12 for Protein and 30 for UAV).

The parameter generator in Algorithm 1, line 5 generates a new instance of the selected function class by randomly sampling shift vectors and rotation matrices. These instance-specific parameters constitute θ\theta. Our implementation follows the public instance-generation routines in https://github.com/MetaEvo/Symbol/blob/main/dataset/generate_dataset.py#L154-L167 and https://github.com/MetaEvo/MetaBox/blob/5565a281ce4e03efc1f6a41230507d5b66637ace/src/environment/problem/SOO/COCO_BBOB/bbob_numpy.py.

We have revised our paper in Section 5.1 and Appendix C.2 to clearly specify the benchmark classes, problem dimensions, evaluation splits, and parameter generation details. Thank you very much.

[1] Symbol: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning. ICLR 2024.

[2] MetaBox-v2: A unified benchmark platform for meta-black-box optimization. NeurIPS 2025.

Q2 Motivation of adding a small constant to softmax

Thank you for pointing this out. The constant ϵ\epsilon is intended to provide a minimum sampling probability for every function class, reducing the risk that some classes are rarely sampled. You are correct that if ϵ=0.05\epsilon=0.05, KK cannot exceed 20. After double checking our implementation, we find that the written-down setting of ϵ=0.05\epsilon=0.05 is a typo. Specifically, we set ϵ=0.05\epsilon=0.05 for CEC training, which consists of 10 classes, and ϵ=0.02\epsilon=0.02 for BBOB training, which consists of 24 classes. A requirement for ϵ\epsilon is 0<ϵ<1/K0<\epsilon<1/K to ensure positive probabilities. It is also interesting for future work to investigate the dynamic configuration of ϵ\epsilon.

We have revised Section 4.1 (Eq. 6) and Appendix C.1 (Table 2) to clarify this setting. Thank you again for pointing this out!

W3 Unnumbered equations & forward reference to Figure 1(b)

Thank you for pointing this out! We have numbered all display equations (Eq. (1)–Eq. (9)) throughout the paper and added an explicit forward reference to Figure 1(b) in Section 1 alongside Figure 1(a) to improve readability.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

Thanks for your detailed response and additional experimental results. At the middle of the figure (https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/teacher_random_3panel.png), HATS with Random Search teacher and HATS with MadDE teacher achieve similar final performance. Obviously, Random Search and MadDE have significantly different performance. So it raises a new question, if they could learn similar behaviours or achieve similar performance by learning from different teachers, what does HATS actually learned form the teachers? Or why sometimes the teacher selections make difference (i.e., CMA-ES vs PSO vs MadDE), sometimes not (i.e., MadDE vs Random Search)?

AuthorsOpenReview ↗

Follow-up response by the authors

Thanks for your follow-up question! However, we would like to first clarify the teacher optimizer (or reference optimizer, more precisely) used in HATS. In our paper, it only provides a reference metric to score instances (Eq. (4)), rather than directly training the policy via imitation or distillation. Policy training is determined entirely by the MetaBBO backbone's own learning objective. We have already used the term "reference optimizer" in the revision to make this clearer.

We introduce this reference simply because raw objective values across instances differ in scale and solvability. An instance-specific baseline is therefore needed to prioritize tasks near the agent's current learning frontier. Thus, as emphasized in our response to W1 & Q1, the reference only decides which instances the agent trains on, while the backbone itself determines what and how well the agent learns.

Following this logic, the middle panel shows that the two references do affect learning differently. MadDE leads to much faster improvement in the middle stage, although both eventually reach a similar level. This similar final level is expected because both runs train the same MetaBBO backbone, which determines the eventual performance given sufficient updates. Similar final scores also do not imply similar learned behaviors. Reference choice simply shapes curriculum quality rather than setting a performance ceiling. Consistent with our response to W1 & Q1, a stronger reference is not necessarily better; one matching the agent's capability range (e.g., MadDE) works best. A much weaker reference (Random Search) makes task prioritization less informative and slows convergence. Conversely, an overly strong reference (e.g., CMA-ES) makes too many tasks look overly hard, hurting the curriculum and final generalization.

We sincerely thank you for taking the time to review our work. Your constructive feedback has greatly helped us improve the clarity of our paper. We hope our clarifications and revisions have addressed your concerns. Please let us know if you have any further questions :)

Reviewer_9Y8dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies dynamic training task selection for Meta-Black-Box Optimization (MetaBBO). The authors first observe a weak correlation between instance diversity in an ELA-based latent space and training value estimated through leave-one-out evaluation. Motivated by this observation, they propose Hierarchical Active Task Selection (HATS), which measures task utility using the performance gap between the current optimizer and a teacher algorithm, instantiated as MadDE. HATS combines bandit-based function-class selection with instance replay based on utility and staleness. Experiments across multiple MetaBBO backbones, synthetic benchmarks, and application tasks show improvements over static training and domain randomization in most evaluated settings.

Critical strengths

1.The paper addresses a meaningful problem: diverse training instances are not necessarily useful for the current MetaBBO policy. Dynamically adapting task selection to the optimizer’s capabilities is well motivated.

2.Experiments cover multiple backbones and synthetic and application tasks, with encouraging improvements. In Table 1, HATS reduces the average normalized objective from 0.163 to 0.065 over DR with CEC-based training. The component ablations further support the framework’s effectiveness.

Critical weaknesses

1.The weak diversity–LOO correlation does not establish that the teacher–student gap predicts learning gains; this gap may also reflect teacher bias or student capacity limitations. The MAB/UAR ablations do not isolate the utility signal’s effectiveness.

2.Comparisons mainly against static training and DR do not sufficiently distinguish HATS from existing prioritized replay approaches. Although component ablations are provided, comparisons with compatible PLR-style methods under matched settings would clarify the benefits of teacher-based scoring and hierarchical scheduling.

Questions for authors

1.Can the authors provide evidence that UU predicts actual learning gains, for example by relating utility at a fixed policy checkpoint to validation improvement after a controlled number of updates? How does this relationship compare with diversity-based or random task scoring?

2.How does HATS compare with a compatible PLR implementation or a flat prioritized replay baseline using the same teacher-based utility?

6 comments
AuthorsOpenReview ↗

Responses by the authors (1/2)

Thanks for your valuable and encouraging comments. Below please find our responses.

W1 & Q1 Relationship between LOO-quality and proxy utility

Thanks for your insightful comments! We agree that the weak diversity–LOO correlation alone does not establish the predictive value of our proposed proxy utility. Our motivation is to use the teacher–student gap as a task-specific reference for potential improvement, and it is worth validating whether this potential translates into actual learning gains.

We conduct an illustrative experiment to examine this. Following your suggestions, we visualize the relationship between different metrics against the LOO quality in https://anonymous.4open.science/r/hats-rebuttal-figures-8E48/loo_validation_3panel.pdf. We start from a fixed mid-training checkpoint, randomly generate 32 instances, and compare three metrics: 1)our proposed teacher-student gap utility in Eq. (4), 2) the landscape latent diversity in Eq. (2), and 3) random score. Here we can observe that our proposed utility exhibits a stronger correlation with LOO quality (Pearson r=0.734r=0.734, Spearman ρ=0.675\rho=0.675) than latent diversity (r=0.184r=0.184, ρ=0.222\rho=0.222). Its Spearman correlation also exceeds the central 95% range of correlations obtained from 200,000 random scores ([−0.350,0.349][-0.350, 0.349]), whereas the diversity correlation falls within this range. This provides clear evidence that our proposed teacher-student performance gap utility offers a better predictive signal towards the instances' quality than the diversity-based scoring.

We have added this experiment to Section 5.2 (RQ5, Figure 5). Thank you very much.

AuthorsOpenReview ↗

Response by the authors (2/2)

W2 & Q2 Comparisons with compatible PLR-style methods

Thanks for your constructive questions! This paper studies active data scheduling within MetaBBO training. In this area, most prior work still relies on static training or DR-style augmentation, and recent work such as Diverse-BBO [1] remains essentially static rather than actively adjusting the training distribution. To the best of our knowledge, HATS is a first step toward active data generation in MetaBBO.

We agree that comparisons with more baselines would clarify the benefits of HATS. Following your suggestions, we add three comparisons with compatible PLR-style methods using SYMBOL [2] on both CEC and BBOB training sets:

  • Flat replay: retains our utility but removes function-class-level hierarchical scheduling.
  • PLR⊥-L1: combines the L1 value-loss ∣R^t−V(st)∣|\hat{R}_t-V(s_t)| scoring proposed in [3] with the replay-only training updates of PLR⊥ [4].
  • HATS-L1: retains our hierarchy but replaces teacher-gap scoring with L1.

The L1 loss measures the critic's prediction error, providing an alternative learning-potential signal to our teacher-referenced performance gap. We compute it using SYMBOL's native n-step return targets while retaining its training objective. Experimental results can be found in the following table.

Base functionsTraining strategyCECBBOBProtein DockingUAVAvg. ScoreAvg. Rank
MA-BBOBStatic0.245 ± 0.2020.237 ± 0.1890.636 ± 0.1080.432 ± 0.2610.38712.25
Diverse-BBOStatic0.302 ± 0.1950.359 ± 0.1560.295 ± 0.0710.484 ± 0.2620.3611.25
BBOBStatic0.251 ± 0.1770.262 ± 0.1340.395 ± 0.1090.486 ± 0.2800.34912.5
BBOBDR0.126 ± 0.1080.149 ± 0.1070.163 ± 0.0560.356 ± 0.2360.1997
BBOBFlat replay0.088 ± 0.0810.090 ± 0.0860.415 ± 0.0940.367 ± 0.2180.247.25
BBOBPLR⊥-L10.118 ± 0.0970.113 ± 0.0790.367 ± 0.0870.226 ± 0.1930.2065.5
BBOBHATS-L10.228 ± 0.1840.215 ± 0.1530.762 ± 0.1200.239 ± 0.2500.36110
BBOBHATS (Ours)0.076 ± 0.0750.110 ± 0.0900.349 ± 0.0840.219 ± 0.2140.1893.75
CECStatic0.181 ± 0.1320.175 ± 0.1140.513 ± 0.1230.308 ± 0.2530.29410
CECDR0.125 ± 0.1080.162 ± 0.1330.093 ± 0.0480.276 ± 0.2700.1645.75
CECFlat replay0.130 ± 0.1140.160 ± 0.1190.383 ± 0.0910.328 ± 0.2110.258.75
CECPLR⊥-L10.081 ± 0.0840.127 ± 0.1090.360 ± 0.0960.261 ± 0.1940.2075.5
CECHATS-L10.070 ± 0.0730.097 ± 0.0870.345 ± 0.0790.285 ± 0.2030.1994.25
CECHATS (Ours)0.043 ± 0.0510.064 ± 0.0590.095 ± 0.0500.102 ± 0.2250.0761.25

Here we observe that HATS achieves the best average score and rank within both training groups, supporting the combined benefits of teacher-based utility and hierarchical scheduling.

We have added these three PLR-style baselines into Section 5.1, Table 1, and Section 5.2 (RQ1) to strengthen our discussion. Thank you again for your supportive feedback!

[1] Instance Generation for Meta-Black-Box Optimization through Latent Space Reverse Engineering. AAAI 2026.

[2] Symbol: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning. ICLR 2024.

[3] Prioritized Level Replay. ICML 2021.

[4] Replay-Guided Adversarial Environment Design. NeurIPS 2021.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

Thank you for the thoughtful and thorough response. I particularly appreciate the additional utility-validation experiments, including the controlled continuation study, and the compatible replay baselines. These additions substantially address my original concerns and strengthen the empirical support for HATS.

I would appreciate your interpretation of an interesting pattern in Table 1 concerning hierarchical scheduling and the teacher–student performance-gap utility in Eq. (4), which scores tasks using the performance difference between the current learned optimizer and the reference optimizer, MadDE.

When BBOB provides the base function classes for training, hierarchical scheduling improves the average score across the four evaluation domains with this utility (Flat replay: 0.240 → HATS: 0.189). However, with L1 value-prediction-error scoring, the hierarchical version performs worse (PLR⊥-L1: 0.206 → HATS-L1: 0.361). When CEC provides the base functions, hierarchical scheduling improves the reported average score under both scoring methods, although the improvement is much larger with the teacher–student performance-gap utility (0.250 → 0.076) than with L1 scoring (0.207 → 0.199).

Could you share your interpretation of why this pattern differs between the two base-function sources? In particular, might the two scoring methods differ in their suitability for allocating training budgets across function classes, beyond prioritizing individual instances?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and Area Chair,

Thank you for your careful reviews and constructive discussions. In the revised PDF, all updates are highlighted in red.

Direct validation of the regret-gap utility (9Y8d, rLAY). Evaluating generated instances from a mid-training SYMBOL checkpoint, our reference–agent log-gap utility correlates much more strongly with leave-one-out (LOO) training quality than landscape latent diversity and clearly exceeds random-ranking baselines (Section 5.2, Figure 5). Controlled continuation runs across easy, moderate, and overly hard utility zones further confirm that prioritizing moderate-gap instances consistently yields the largest test gains, whereas overly hard tasks degrade mid- and late-stage policies (Appendix D.1, Table 3).

Compatible prioritized replay baselines and class-coverage analysis (9Y8d). On both BBOB and CEC, HATS consistently outperforms compatible PLR-style baselines (Flat replay, PLR$^{\perp}$-L1, and HATS-L1) across synthetic and real-world test suites, achieving the best overall score and average rank in both training groups (Sections 5.1-5.2, Table 1). Training-log diagnostics further show that hierarchical scheduling protects balanced class-level update coverage, whereas replacing our reference-based gap with L1 value-loss error leads to late-stage critic instability on heterogeneous function classes (Appendix D.8, Table 5, Figures 12-13).

Reference optimizer sensitivity and post-surpassing dynamics (M8sN). Evaluating CMA-ES, PSO, and MadDE across multiple MetaBBO backbones shows that HATS remains robust to reference choice, while capability-matched MadDE consistently achieves the best overall rank (Appendix D.2, Table 4). Replacing MadDE with Random Search mid-training further confirms that the reference only scores instances rather than capping the learner: although mid-stage convergence slows once the gap turns negative, test performance continues improving steadily without bandit entropy collapse (Appendix D.2, Figure 6).

Matched function-evaluation budgets (rLAY). Accounting for both learner rollouts and reference evaluations, SYMBOL-HATS introduces no extra evaluation cost by reusing SYMBOL's native reference trajectories (Appendix D.7). On teacher-free backbones (LDE and B2Opt), under strictly matched total training evaluations on CEC, HATS still consistently outperforms DR and Static training (Section 5.2, Appendix D.7, Figure 11).

Other revisions. Throughout the manuscript and figures, we unified the term "teacher" to "reference optimizer" to clarify that it provides a task-selection reference metric rather than policy distillation (Abstract, Sections 1, 3.2, 5, and 6, Appendices C-D). Section 4.1 and Appendix C.1 (Table 2) clarify the minimum class-sampling probability in Softmax, while Section 5.1 and Appendix C.2 specify benchmark classes, problem dimensions, evaluation splits, and parameter generation. We also numbered all display equations and added explicit figure cross-references.

Thank you again for your insightful comments, which have helped us substantially improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.