acceptodds
Under review as a conference paper at ICLR 2027

When Not to Trust a Global Gaussian Process Posterior: Local Scoring and Progressive Dissonance for Multi-Modal Scientific Discovery

Abstract

Gaussian-process Bayesian optimization (GP-BO) is a standard approach to budget-constrained black-box optimization, but its acquisition decisions rely on the quality of a global GP posterior. We study when this reliance becomes unreliable in scientific discovery problems with phase boundaries, feasibility gaps, threshold-derived responses, and disconnected high-value regions. In these settings, a globally smooth posterior can be miscalibrated precisely in high-response regions: on LGPS, nominal % GP coverage drops from overall to in the top response decile. We introduce two alternatives that reduce dependence on global posterior ranking: ProSe, which combines local -nearest-neighbor scoring with spatial anti-clustering, and ATLAS, which searches for progressive disagreement with a hierarchy of simple models while probing spatial frontiers. Our results show a regime split rather than universal superiority. On a smooth synthetic control, several GP methods and random search attain Recall@100 — the fraction of distinct target peaks recovered within 100 evaluations — of . In contrast, on a strict LGPS compositional-family holdout, ATLAS achieves Recall@50/100 of and reaches successful discoveries in a median of evaluations versus for MAP-Elites, though its advantage over MAP-Elites is directional and borderline after Holm correction (). On discriminative steel search, ATLAS and SAGE — a -NN-scored baseline that anchors part of its candidates on observed points — each significantly outperform eight of ten alternative GP-based acquisition, quality-diversity, and randomized-search baselines after Holm correction. Finally, an audit of five look-up benchmarks constructed from public materials datasets finds that four show no separation among methods under our protocol, due to saturation, positional-encoding collapse, or random-search matching. These results suggest using global posterior acquisition on smooth, calibrated landscapes, and favoring locality, dissonance, and diversity when its modeling assumptions are violated.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian optimization, black-box optimization, Gaussian processes, model misspecification, sequential experimental design, materials discovery, quality-diversity, benchmark validity, compositional data

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dEmnOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces three new Bayesian optimization methods designed to perform competitively on datasets where use of global GP posterior acquisition is problematic. The performance of these and several other Bayesian optimization methods is explored on a range of real-world and synthetic black-box optimization problems. The results show competitive performance of the newly proposed ATLAS/SAGE methods on one of the datasets outperforming 8 out of 10 alternative methods. The analysis provides useful insights into when alternatives to GP posterior acquisition, such as the approaches adopted by the new methods, can be advantageous. A key insight is that global posterior acquisition is appropriate on smooth calibrated landscapes but methods exploiting locality, dissonance and diversity should be used when those modeling assumptions are violated. The paper also identifies deficiencies in some of the public datasets, calling into question their usefulness for evaluating Bayesian optimization techniques,

Critical strengths

The paper is clearly written. The experiments cover an extensive range of methods and datasets with much additional experimental detail in the appendices and provide some useful insights. The designs of the new methods seem heuristically reasonable. An objective and honest account of the performance of the newly introduced methods is provided.

Critical weaknesses

The paper is addressing two things at once:

(a) introducing, explaining and showcasing the new Bayesian optimization methods
(b) understanding how performance of these and other Bayesian optimization methodologies relate to data landscape and identifying deficiencies in some of the benchmark datasets.

Perhaps as a result of covering both aspects at once the central messages of the paper have become somewhat mixed-up/blurred

A more rigorous showcasing of the new methods would need to include: expanded justification of the construction and choice of components in the models e.g. through more extensive ablation studies, justifying parameter/weight choices together with sensitivity analysis etc. The reader would also expect the new methods to be shown competitive against others on several relevant datasets. The paper in its current form doesn't achieve this combination.

If a primary aim is a scientific understanding of how methodologies need to match dataset landscape then the simultaneous introduction of these new methods may be seen as more confusing than helpful. The methods are quite complex and as yet not fully understood making scientific diagnostic conclusions from them difficult.

I feel that the paper in its current form has failed to cover both (a) and (b) adequately (particularly (a)). At least in part this may be due to the challenge of attempting to cover both at once in a short conference paper.

Questions for authors

(1) Can you please clarify whether ProSe, SAGE and ATLAS are intended for diagnostic purposes or intended to be competitive optimization methods.

(2) If intended as competitive optimization methods can you provide some more detail on how you went about designing the new methods and to what extent they've yet been tuned or subjected to ablation studies.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dEmn

Revision note. We revised the title to “Recall-Oriented Search on Retrospective Materials Emulators” and updated the abstract to reflect the narrower claims and additional evidence. The original TL;DR no longer reflects the revised conclusions; please refer to the revised abstract and manuscript.

Thank you for asking us to clarify the contribution and its relationship to competitive optimization. We now ask a narrower empirical question: how do specified search policies behave on target-region retrieval and best observed response in fixed retrospective materials emulators? We do not claim a general GP-failure diagnostic, a regret guarantee, or universal superiority.

ProSe combines local value estimation with anti-clustering; SAGE is a separately parameterized anchored reference; ATLAS combines residual-guided proposals, coverage/dissonance scores and a model hierarchy. The revised methods specify the scoring, geometry, proposal allocation and transitions. These combine known ingredients, and their coefficients were hand-designed with historical tasks in view; we do not claim each ingredient is new or that broad tuning robustness has been demonstrated.

The quantitative limits are central to the revision: ProSe records 0/100 retrievals on each primary task, and ATLAS versus MAP-Elites is not significant after the primary Holm28 correction. Only two of ten specified ATLAS component/sensitivity contrasts support corrected benefits. The later joint residual-pathway deletion was selected after earlier results and changes several pathways at once; its results do not establish the necessity of any individual component. The available weight controls and post-hoc target/radius re-scoring do not establish general sensitivity robustness.

An additional exploratory band-gap emulator likewise gives ATLAS/MAP target hit rates of .36/.28 without a corrected significant difference, while the four tested GP variants achieve .88–.99. The mean best observed normalized responses are .887719 for ATLAS and .861510 for MAP; a corrected sign test favors ATLAS in direction, and the GP variants are also favored over both. The sign test does not test the mean difference itself. The previously studied subset includes the target points in the emulator support. We therefore report this counterevidence and restrict conclusions to the specified retrospective tasks, without claiming prospective discovery or generalization to unseen materials.

Thank you for responding to my review. I'm having difficulty seeing how my concerns are being addressed as you've not adopted the usual rebuttal convention of addressing each of my points individually and in turn. It will be challenging for me to look at the changed emphasis plus substantial paper-rewrite within the remaining time-limit but I will do my best. Meanwhile I maintain my score.

AuthorsOpenReview ↗

Point-by-point clarification

Thank you for explaining that our response was hard to follow. We apologize. Here are direct answers to your two questions and the related concerns:

  1. Purpose. ProSe and ATLAS are proposed as search policies to be evaluated against established methods; SAGE is a separately parameterized reference. They are not presented as a diagnostic for GP failure. We now ask how these policies perform on target-region retrieval and best observed response on fixed retrospective emulators (Sections 1 and 3).

  2. Design and ablation. The methods and hand-chosen weights are specified in Sections 3.2–3.4. Section 4.2 lists the fixed controls, and Section 5.2 reports their results: only two of ten ATLAS component/sensitivity comparisons support a benefit after correction. They do not establish that individual components are necessary. The weights were developed with historical tasks in view; broad tuning robustness is not established.

  3. Competitiveness across datasets. On the two primary tasks, ATLAS’s observed retrieval was higher than MAP-Elites’, but neither difference was significant after multiple-comparison correction. ProSe retrieved neither target in 100 runs per task. On the exploratory band-gap emulator, the tested GP variants had higher target retrieval than ATLAS and MAP-Elites. These fixed retrospective tasks do not establish general superiority or performance on unseen materials (Sections 5 and 6).

We have also rewritten the conclusion in Section 8 to state the question, main findings, and limits more plainly. We appreciate your taking the time to review the revision.

I have looked quickly at the revised manuscript in an attempt to comment usefully before final discussion closure. At the moment I think I will at best be able to increase my rating to 2 (weak reject) as I feel more has to be done to justify the value of the new methods and to show how the parameters and combination of components were arrived at (as per my earlier comments on ablation etc). I will look some more after discussion-closure to see whether I can justify raising this to 2 or whether I retain my 1 marking.

Reviewer_8zC3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses the limitations of global GP posteriors in multi-modal scientific optimization problems. The authors propose two methods, ProSe and ATLAS, which use local scoring, spatial exploration, and progressive model disagreement to reduce reliance on global GP predictions. Experiments on synthetic and real-world materials datasets demonstrate improvements in certain challenging settings, while also showing the limitations of existing benchmarks and the importance of posterior calibration.

Critical strengths
  1. Important Research Problem: The paper addresses the limitations of global GP posteriors in multi-modal and irregular scientific optimization landscapes.

  2. Interesting Methodological Perspective: The proposed ProSe and ATLAS methods explore alternative strategies based on local scoring, model disagreement, and spatial diversity.

  3. Extensive Experimental Evaluation: The authors conduct experiments on synthetic functions and multiple real-world materials datasets, comparing against a broad range of baselines.

  4. Careful Statistical Analysis: The paper includes repeated experiments, statistical significance tests, and multiple-comparison corrections to evaluate the reliability of its findings.

  5. Transparent Discussion of Limitations: The authors acknowledge inconsistent performance, benchmark saturation, and reproducibility issues, providing a balanced interpretation of the experimental results.

Critical weaknesses
  1. Limited Methodological Novelty: The proposed methods primarily combine existing techniques, including k-NN scoring, clustering, model disagreement, and spatial exploration, with several heuristic design choices.

  2. Inconsistent Experimental Performance: Neither ProSe nor ATLAS consistently outperforms existing baselines. ProSe performs poorly on several key tasks, while ATLAS is outperformed by MAP-Elites in some settings.

  3. Limited Statistical Significance: The improvements of ATLAS over MAP-Elites on the LGPS family holdout are not statistically significant after Holm correction, weakening the evidence for its superiority.

  4. Benchmark Reliability: Four of the five public materials benchmarks fail to meaningfully distinguish between methods due to saturation or representation issues, limiting the strength of the empirical conclusions.

  5. Insufficient Comparison with Adaptive Surrogates: The paper focuses primarily on limitations of global GP models but does not sufficiently investigate whether non-stationary or locally adaptive GP surrogates could address these issues without abandoning GP-based optimization.

Questions for authors
  1. The proposed methods rely on several existing techniques and heuristic design choices. Could the authors clarify the key methodological contributions beyond combining these components? More comprehensive ablation studies demonstrating the necessity of each component would strengthen the paper.

  2. The paper shows the limitations of global GP posteriors, but it remains unclear whether these issues could be addressed by more flexible surrogate models. Could the authors include comparisons with non-stationary or locally adaptive GP approaches? Demonstrating consistent improvements over such methods would also strengthen the central argument.

  3. The improvements of ATLAS over MAP-Elites are not statistically significant after Holm correction, and some results vary across platforms. Could the authors investigate these discrepancies and provide stronger evidence of consistent improvements? In particular, statistically significant gains across multiple independent benchmarks and reproducible results would substantially increase my confidence in the proposed approach and make me more inclined to raise my score.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 8zC3

Revision note. We revised the title to “Recall-Oriented Search on Retrospective Materials Emulators” and updated the abstract to reflect the narrower claims and additional evidence. The original TL;DR no longer reflects the revised conclusions; please refer to the revised abstract and manuscript.

Thank you for highlighting the platform discrepancy, limited component evidence, and lack of consistent significant gains. We agree that our results do not establish general superiority. The earlier public lookup studies included saturation and representation limitations, and the revised primary argument does not treat them as independent confirming benchmarks. On the revised fixed primary tasks, ATLAS versus MAP-Elites retrieval was .515 versus .435 for LGPS and .635 versus .530 for steel (200 paired repetitions per task). The paired effects were .080 (unadjusted 95% bootstrap interval [−.005, .170], Holm28 p=1.000) and .105 ([.020, .190], Holm28 p=.374569). Neither test establishes an ordering or equivalence.

We also examined a previously studied 83-composition band-gap subset under a retrospective three-neighbor emulator. ATLAS/MAP target hit rates were .36/.28 (paired effect .08, unadjusted 95% interval [−.04, .20], Holm26 p=1.0); four GP variants scored .88–.99 and each was favored over both policies in the corrected retrieval comparisons. ATLAS had a higher mean best normalized response than MAP (.887719 versus .861510), with the paired sign test favoring ATLAS after correction; all four GPs also favored higher best response over both policies. The sign test concerns directions, not the mean difference itself. Six target points form one target region, are excluded from initialization, and remain in the emulator support. This is an exploratory, previously studied virtual task, not a blind independent confirmation or physical discovery.

For component evidence, only two of ten specified ATLAS component/sensitivity contrasts support corrected benefits of the full policy: retaining the combined frontier/density coverage score on LGPS, and full progression versus a constant-model policy on steel. A later joint deletion of residual-dependent pathways was designed after inspection of previous GP results and cannot isolate an individual component or establish necessity. We report all contrasts and retain the limits on sensitivity to the hand-designed choices.

An exploratory continuous GP-UCB run with a previously selected length scale attained .43 retrieval at budget 100 over 100 LGPS repetitions. The saved results were independently re-aggregated, but the optimizer was not rerun. Its initial designs match those reconstructed using the primary seed rule, while the separate GP-EI supplement uses a different seed namespace; case-file identity remains unresolved. It is a descriptive single-task comparison, not an independent benchmark or a paired significance result.

The revised LGPS task aggregates duplicate compositions with deterministic neighbor ordering. An earlier matched-input Linux/macOS oracle witness agreed within about 2.2×10⁻¹⁶, but a full cross-platform optimizer rerun has not been established. Separately, same-machine, same-code, same-seed reruns reproduced 200 local-candidate GP and 400 matched ATLAS/MAP trajectories and metrics; this does not replicate the original 200-repetition primary campaign, test a different implementation, or add independent materials. The tested local-candidate GP is a global Matérn model with a locally sampled candidate pool, not published TuRBO or an explicitly nonstationary GP. Numerical warnings remain, and adaptive/nonstationary GP alternatives are not exhaustively tested.

Finally, the LGPS family is an operational composition rule containing two targets; standard Li₁₀GeP₂S₁₂ is outside it, and the XRD peak-height proxy is not calibrated to phase fraction or conductivity. The post-hoc radius sweep re-scores saved trajectories without new optimization or hypothesis tests. It shows sensitivity to target/radius choices, not physical phase discovery or a confirmed method ordering.

Reviewer_dU2LOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

 

The authors introduce two methods, ProSe and ATLAS with the goal of modelling and performing Bayesian optimization over irregular objective functions commonly found in scientific applications. Although the idea is very topical and the paper is well-written, there are a number of points I would like to see addressed before I would fell comfortable recommending acceptance as described below.

 

Critical strengths

 

The topic of the paper is clearly relevant. Nonsmooth functions are commonplace in the sciences. Bayesian optimization is a powerful methodology for performing black-box optimization in low-data regimes common to the experimental sciences, and current Gaussian process machinery struggles to model irregular optimization landscapes.

 

Critical weaknesses

 

I demarcate between major points and minor points below.

 

MAJOR POINTS

 

  1. The main point I would like to see addressed is displaying Bayesian optimization traces in the main paper. Regret curves (or in the case where regret is not available "best so far" or similar) would allow for clear visual distinction against baselines.

  2. The choice of baselines is somewhat mismatched with the application e.g., TuRBO is a high-dimensional optimizer and ROBOT seeks diversity in the solutions. For nonsmooth functions, HEBO (which the authors cite) is a relevant baseline since it features both input and output warping by Beta-Kumaraswamy input warping [8] and Yeo-Johnson and Box-Cox transforms to address nonstationarity and heteroscedasticity respectively. I would also recommend looking into [7] which also considers nonsmooth functions. It may also be worth looking into the heteroscedastic Bayesian optimization literature for more nuanced discussion as to whether nonsmoothness is caused by nonstationarity in the underlying objective or heteroscedasticity in the noise process.

  3. It would be great if the authors could release their codebase for reproducibility of the empirical results.

 

MINOR POINTS

 

  1. The reference format is inconsistent in so far as the paper title is provided for some references but not others. It would be great if this was fixed.

  2. When introducing Bayesian optimization in the introduction the authors should cite the originating papers for the method [1, 2] as discussed in [3].

  3. The originating papers for Thompson sampling [4] and expected improvement [5] should be cited when introducing the acquisition functions. The originating paper for expected improvement is discussed in [3].

  4. In the related work it would be worth citing [6] as a reference work on applications of Gaussian processes in scientific domains.

 

REFERENCES

 

[1] H.J. Kushner (1962). A Versatile Stochastic Model of a Function of Unknown and Time Varying Form. Journal of Mathematical Analysis and Applications 5(1):150–167.

[2] H.J. Kushner (1964). A New Method of Locating the Maximum Point of an Arbitrary Multipeak Curve in the Presence of Noise. Journal of Basic Engineering 86(1):97–106.

[3] Garnett, R., Bayesian optimization. Cambridge University Press. 2023.

[4] Thompson, W. R., 1933. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4), pp.285-294.

[5] V. R. Saltenis (1971). One Method of Multiextremum Optimization. Avtomatika i Vychislitel’naya Tekhnika (Automatic Control and Computer Sciences) 5(3):33–38.

[6] Griffiths, R.R., 2022. Applications of Gaussian processes at extreme lengthscales: From molecules to black holes. University of Cambridge (United Kingdom).

[7] Bodin, E., Kaiser, M., Kazlauskaite, I., Dai, Z., Campbell, N. and Ek, C.H., 2020, November. Modulating surrogates for Bayesian optimization. In International Conference on Machine Learning (pp. 970-979). PMLR.

[8] Snoek, J., Swersky, K., Zemel, R. and Adams, R., 2014, June. Input warping for Bayesian optimization of non-stationary functions. In International conference on machine learning (pp. 1674-1682). PMLR.

 

Questions for authors

 

Questions included in the weaknesses section.

 

4 comments
AuthorsOpenReview ↗

Response to Reviewer dU2L

Revision note. We revised the title to “Recall-Oriented Search on Retrospective Materials Emulators” and updated the abstract to reflect the narrower claims and additional evidence. The original TL;DR no longer reflects the revised conclusions; please refer to the revised abstract and manuscript.

Thank you for requesting optimization trajectories, stronger GP controls, reproducibility, and the missing historical references. Figure 1 reports target-region retrieval and best observed response over the evaluation budget. They measure different objectives; on noisy steel, best observed response is not latent regret. We have narrowed the framing and no longer infer GP miscalibration from retrieval differences.

The supplemental study compares four GP-EI settings: MLE or grid-MAP fitting, each with pool or continuous acquisition optimization. Continuous EI improves best observed response in all four matched contrasts after Holm correction, while the corresponding retrieval contrasts do not establish a corrected benefit. The tested GPs retrieve less than ATLAS/MAP-Elites on the two primary tasks, but continuous EI achieves higher best observed response on steel. This is a task- and objective-dependent result, not a general GP ordering. The related-work discussion now distinguishes nonstationarity from input-dependent noise and includes the requested historical and flexible-surrogate references.

The original default-HEBO comparison remains on numerical-health hold: selected witnesses have abandoned fits or prediction fallbacks, and their frequency in the original campaign is unknown. The separate HEBO-Adam runs completed the specified checks, but 172 of 200 emitted variance-floor warnings. The tested default backend has no input warp; our separate explicitly warped GPyGP arm uses Kumaraswamy input warping, not a beta-CDF warp. These arms use different GP backends and fitting procedures, so they are not an isolated warping ablation. Both use scalar likelihood noise after response transformation, not an explicitly heteroscedastic noise model. The related-work discussion includes Bodin et al.'s modulating surrogates, which we did not test. A further exploratory local-candidate GP fits a global Matérn model; it is neither published TuRBO nor an explicitly nonstationary GP. Its conclusions are limited by fitting warnings, the fixed datasets and a post-outcome-selected comparison.

An additional exploratory LGPS GP-UCB run with a length scale selected in an earlier LGPS screen achieved retrieval .43 at budget 100 in 100 repetitions. Its saved results were independently re-aggregated; the optimizer was not independently rerun. Its initial designs match those reconstructed from the primary campaign seed rule, whereas the separate GP-EI supplemental comparison uses a different seed namespace. The case-file identifiers differ and their byte-level relationship is unresolved. This is a descriptive single-task result, not a paired significance test or an independent materials benchmark.

On a previously studied 83-composition band-gap subset, four additional GP variants achieved target hit rates .88–.99 versus .36 for ATLAS and .28 for MAP-Elites; all eight GP–policy retrieval comparisons favored the GP after Holm26. This is a retrospective three-neighbor emulator: the six target points remain in its support. MLE variants produced convergence warnings, with fallbacks in the continuous variant. It is neither blind unseen-data validation nor physical discovery.

We repeated the 200 local-candidate GP trajectories and 400 matched ATLAS/MAP trajectories on the same machine with the same code, inputs, packages and seeds, recovering identical recorded trajectories and metrics. Separate code replayed the saved observations and checked the statistics. This is not an independently implemented optimizer or a cross-platform rerun; external code and restricted-input availability remain unresolved, and we do not claim a public release.

The revised bibliography includes the requested references to Kushner (1962, 1964), Garnett, Thompson, Šaltenis, Griffiths, Bodin et al. and Snoek et al. The original Šaltenis article was not retrieved; its identity was corroborated using Garnett's bibliography.

Reviewer_dU2LOpenReview ↗

Clarity Seems to have Deteriorated

 

While I would like to thank the authors for their response, large tracts of the rebuttal appear to have been written by an LLM which has rendered communication nebulous and opaque e.g.,

 

"The original default-HEBO comparison remains on numerical-health hold: selected witnesses have abandoned fits or prediction fallbacks, and their frequency in the original campaign is unknown. The separate HEBO-Adam runs completed the specified checks, but 172 of 200 emitted variance-floor warnings. The tested default backend has no input warp; our separate explicitly warped GPyGP arm uses Kumaraswamy input warping, not a beta-CDF warp. These arms use different GP backends and fitting procedures, so they are not an isolated warping ablation. Both use scalar likelihood noise after response transformation, not an explicitly heteroscedastic noise model. The related-work discussion includes Bodin et al.'s modulating surrogates, which we did not test. A further exploratory local-candidate GP fits a global Matérn model; it is neither published TuRBO nor an explicitly nonstationary GP. Its conclusions are limited by fitting warnings, the fixed datasets and a post-outcome-selected comparison."

 

What does numerical-health hold mean? I have never seen this term in the optimization literature or otherwise. What are selected witnesses? The takeaway of this paragraph is incredibly unclear.

 

Next, what do the authors mean with this paragraph,

 

"We repeated the 200 local-candidate GP trajectories and 400 matched ATLAS/MAP trajectories on the same machine with the same code, inputs, packages and seeds, recovering identical recorded trajectories and metrics. Separate code replayed the saved observations and checked the statistics. This is not an independently implemented optimizer or a cross-platform rerun; external code and restricted-input availability remain unresolved, and we do not claim a public release."

 

What point in my rebuttal does this paragraph address? Lastly, the re-written conclusion is more unclear now then it was in the original manuscript.

 

AuthorsOpenReview ↗

Clarification of HEBO, repeatability and conclusion

Thank you for pointing out that our wording was unclear. We apologize. ‘Numerical-health hold’ is our own label, not a standard term: it means that we do not use the original default-HEBO results to claim a fair performance comparison. We reran four selected cases using the original inputs and repetition IDs 0 and 1 on each task. Three failed our checks because some fitting steps were abandoned or prediction fell back to a different procedure. We do not know how often this occurred in the original campaign, and we do not conclude that HEBO is generally unusable. In a separate 200-run experiment using Adam fitting, all runs finished, although 172 produced variance-floor warnings; the warnings alone do not establish their severity.

You asked about releasing the code. Our rerun was only a limited repeatability check: the same machine, code, inputs and seeds reproduced 200 GP and 400 ATLAS/MAP-Elites trajectories. It is not independent reproduction and does not replace public code or input access, which we cannot yet promise. We agree that this does not resolve your request.

Our conclusion is that performance depends on the task and objective. On the two main tasks, the observed ATLAS–MAP-Elites retrieval differences were not statistically significant after correction; on an exploratory band-gap task, the tested GPs retrieved the target more often. These fixed-emulator results do not establish general superiority, GP miscalibration or discovery of new materials. We will clarify the conclusion accordingly.

Reviewer_dU2LOpenReview ↗

Maintaining Score

 

I maintain my score following the rebuttal.

 

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.