acceptodds
Withdrawn from ICLR 2027

GIFT-Evo: Feedback Adaptation through Geometric-Inductive Operators for Low-Budget Black-Box Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned evolutionary optimizers encode cross-task strategies into fixed parameters via offline training, allowing direct search on new problems without per-task tuning for low-budget black-box optimization. However, existing methods do not retain real-time search feedback, preventing the policy from exploiting past outcomes to guide later decisions and accelerate convergence. Online parameter updates adapt to the target but require extensive samples, conflicting with low budgets. To this end, we propose GIFT-Evo, which introduces a lightweight search knowledge channel that accumulates individual acceptance and improvement history. This state modulates offspring generation to better suit the current problem, thereby accelerating convergence, while keeping parameters fixed. Since search knowledge must be translated through the model, and absolute-coordinate-based models fail under scale changes, we further design a geometric-inductive operator framework based on relative population geometry to ensure consistent semantics across problem scales. Experiments on multiple baselines show that GIFT-Evo achieves significant advantages in both low-budget and cross-scale deployment scenarios.

Primary area:
optimization
Keywords:
Black-Box Optimization, Meta-Learning, Low-budget Optimization

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_4G9eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GIFT-Evo, a learning-based evolutionary optimizer designed for black-box optimization under a limited budget of function evaluations. The method combines offline-trained operators with the individual's search state that records information. During iteration, the model parameters remain fixed while the search state continuously evolves. Experimental evaluations include the B2Opt-based BBOB implementation, multi-dimensional deployment, component ablation studies, and validation on two scenarios.

Critical strengths

The method described in the text is clear and detailed, the experimental results and analysis are quite comprehensive.

Critical weaknesses

My main concern is that these surprising and prominent results might stem from the interaction between the construction of the benchmark tests and the update rules that depend on the origin. Table 16 particularly surprised me: under 1000-dimensional conditions, GIFT-Evo achieved nearly optimal performance on many functions after just 3100 function evaluations; in contrast, other algorithms still showed significant gaps when using the original evaluations. Similar large gaps were observed on several other functions. These results have raised my concern: is this method actually learning to locate the unknown global optimum, or is it merely converging toward a particular point that has been specially favored?

This issue is highly similar to Kudela's concern raised in "A critical problem in benchmarking and analysis of evolutionary computation methods" (Nature Machine Intelligence, 2022): when the optimal point of a benchmark function is placed at the center, operators biased toward the center might perform exceptionally well; however, their performance may change significantly when the problem is shifted.

  1. Section 4.1 and Appendix E.3 indicate that the experiment uses an implementation of BBOB paired with B2Opt. In the publicly available B2Opt evaluation script, exps/main.py lines 344–356 explicitly sets xopt to zero for all functions except F05 and F24, and sets fopt to zero during instance generation. Lines 378–391 then loads these instances for evaluation. This directly places the optimal solution x for the majority of functions at the zero point.

  2. F05 provides a noteworthy contrast. According to lines 236-238 of utils.py, the optimal point of this function lies at boundary corner points, not at zero. In Figure 9, when d=100 and all methods use the same 3100 evaluation budget, GIFT-Evo does not demonstrate its significant convergence advantage observed on many other functions, and significantly lags behind in both comparisons. The results for F05 further support the aforementioned concerns.

  3. Having only test functions does not necessarily mean the algorithm is exploiting such properties. Therefore, I carefully examined the specific method definitions of GIFT-Evo. Eq. (7) combines three sets of coefficients to generate the absolute coordinates of candidate solutions.

    Appendix A.4 explicitly states that the sum of these three coefficients does not have to equal 1.

    Let si=cp,i+cc,i+cm,is_i = c_{p,i} + c_{c,i} + c_{m,i}. The update rule can be rewritten as

    xchild,i=sixi+cc,i(xcross,i−xi)+cm,i(xmut,i−xi).x_{\mathrm{child},i} = s_i x_i + c_{c,i}(x_{\mathrm{cross},i} - x_i) + c_{m,i}(x_{\mathrm{mut},i} - x_i).

    When the candidate solutions in each branch are close to one another, this update approximately becomes

    xchild,i≈sixi.x_{\mathrm{child},i} \approx s_i x_i.

    Specifically, by setting cp=ac_p = a and cc=cm=0c_c = c_m = 0, the operator can implement

    xchild,i=axi,0<a<1.x_{\mathrm{child},i} = a x_i, \quad 0 < a < 1.

    For a simple function where the optimal point lies at zero, adopting such a contraction rule is a shortcut to reducing training objectives. The same rule can naturally generalize across dimensions without requiring the model to actually learn the location of unknown optimal points. Even if the coefficient network takes relative values as input, the final update still depends on the absolute coordinates of the candidate solutions.

  4. What particularly puzzles me is that the authors trained only on a few simple functions CECF1 to CECF3 with d=10, and yet report such broad and significant generalization advantages. This result contradicts the trend observed in Lange et al. "Discovering Evolution Strategies via Meta-Black-Box Optimization", which analyzes the relationship between training distribution and generalization performance: their analysis suggests that incorporating more complex, multi-modal, and ill-conditioned training functions helps improve generalization, and that the training process also randomizes the location of optimal solutions.

    In contrast, this paper claims to achieve strong cross-function and cross-dimensional transfer capabilities using only extremely limited and simple training functions, without providing a clear explanation for why such a simple training distribution could support these results. This discrepancy requires a compelling mechanistic explanation and empirical evidence, especially one that rules out the shortcut of shared optimal solution locations between training and test functions.

    Current deployment-phase ablation experiments are insufficient to address this issue, as they fail to distinguish whether the model has learned a search strategy adapted to different target functions or simply a rule that contracts toward the origin based on feedback.

The authors should first clearly define the specific forms of CECF1–CECF3 and the distribution of their optimal solutions, and then verify whether the reported generalization advantages still hold when both training and test functions have randomized optimal solution locations. Additionally, a comparison should be made between the trained model and a simple contraction baseline on both original and shifted instances, recording the sum of fusion coefficients and population trajectories, and evaluating a retrained variant where the sum of fusion coefficients is constrained to 1 but negative weights are allowed.

Without such evidence, I find it difficult to accept that the prominent results achieved under such a simple training setup can be directly attributed to broad, transferable meta-learning capabilities; contraction toward the origin remains a plausible alternative explanation.

Questions for authors
  1. Did the actual BBOB instances used in this paper retain the handling of setting xopt to zero as in B2Opt? Please provide evidence to clearly identify the actual optimal points for each test function. When the optimal points are independently randomly translated, or when using official COCO instances, does GIFT-Evo maintain its advantage under low evaluation budgets, especially its strong performance in high-dimensional cases highlighted in Table 16?

  2. Please report the fusion coefficients of GIFT-Evo on functions F01, F03, and F05, as well as the evolution of the sum si=cp,i+cc,i+cm,is_i = c_{p,i} + c_{c,i} + c_{m,i} with iteration, and analyze whether the update exhibits a continuous trend of converging toward the origin, based on the population center trajectory. Compared to a simple fixed contraction baseline, how significant is GIFT-Evo's advantage on both original and translated instances? If the sum of fusion coefficients is constrained to 1, while retaining negative weights and retraining, do the main performance conclusions still hold?

  3. Please clearly define CECF1–CECF3, specify the instance generation method, and describe the distribution of optimal point locations. How to explain that such remarkable cross-function and cross-dimensional generalization performance can be achieved by training only on three simple functions at d=10, when LES analysis shows that a richer training distribution helps improve generalization?

1 comment
AuthorsOpenReview ↗

For thanks

Thank you for your careful analysis and for highlighting this potential confounding factor. The central objective of GIFT-Evo is to investigate how individual search feedback can guide learned evolutionary operators while keeping model parameters fixed during deployment.

As you noted, origin-related bias is a broader benchmarking concern in evolutionary optimization. Our evaluation follows a B2Opt-derived instance-construction protocol, making your observations directly relevant to the evidence presented in our paper. Although addressing coordinate-origin bias was not the primary focus of this work, ruling out this confound is necessary to substantiate our performance and generalization claims.

In a future revision, we will explicitly document the training functions, instance-generation procedures, and optimum-location distributions. We will incorporate evaluations with independently shifted optima and validated benchmark instances, comparisons against fixed contraction baselines, and a retrained fusion variant whose coefficients sum to one while allowing negative weights. We will also report the fusion coefficients and population trajectories to distinguish origin-dependent contraction from effective search behavior.

We appreciate your suggestions and will revise the framework and conclusions according to the resulting evidence.

Reviewer_US9jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GIFT-Evo combines offline-trained evolutionary operators with per-individual historical feedback states. It reports strong low-budget performance and cross-dimensional deployment, but its incremental contribution over closely related methods remains insufficiently established.

Critical strengths

The method addresses a important problem, i.e., exploiting search feedback without additional function evaluations or test-time parameter updates.

Critical weaknesses
  1. The proposed design largely overlaps with POM and EPOM. Fixed-parameter deployment, normalized fitness features, learned population combinations, and cross-dimensional applicability are already present in POM and EPOM. The contribution appears to be an incremental extension.

  2. POM and EPOM are particularly relevant baselines but are not evaluated. Comparisons with versions augmented by the same feedback state would help determine whether the gains require GIFT-Evo’s architecture or primarily reflect a general history-conditioning benefit.

Questions for authors
  1. How does GIFT-Evo compare with POM and EPOM under matched protocols, including versions equipped with the same feedback state?

  2. Which geometric properties hold for the complete update, and do the gains persist under randomized coordinate shifts, scale changes, and objective offsets?

  3. Do retrained no-state, time-only, and shuffled-history controls confirm that meaningful historical feedback drives the improvement?

1 comment
AuthorsOpenReview ↗

For thanks

Thank you for your thoughtful review and for identifying the comparisons and controls needed to assess GIFT-Evo’s contribution. We will address your concerns in a future revision.

  1. Relationship to POM and EPOM. We will clarify the overlap with these methods and narrow our novelty claims. We will pursue comparisons under matched training and evaluation protocols, including variants equipped with the same feedback state, to distinguish general history-conditioning benefits from contributions specific to our architecture.
  2. Geometric properties of the complete optimizer. We acknowledge that relative input features and the crossover’s local equivariance do not establish translation equivariance of the complete update. We will revise the geometric claims, examine the unrestricted fusion mechanism, and evaluate performance under consistent coordinate translations, scale changes, and objective offsets. Coordinate transformations and objective-value transformations will be analyzed separately.
  3. Contribution of historical feedback. We will strengthen the analysis with independently retrained no-state, time-only, and shuffled-history controls under consistent training conditions and evaluation budgets. These comparisons will separate outcome-dependent feedback from time information and test the importance of associating histories with individual solutions. We will report the findings transparently and revise the framework and conclusions wherever the evidence requires. We appreciate your guidance in strengthening the paper’s empirical foundations.
Reviewer_TR3JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GIFT-Evo, a learned evolutionary optimizer designed for low-budget black-box minimization problems. It combines fixed, offline-trained operator parameters with an exponential moving average state associated with each individual, which encapsulates feedback regarding acceptance, relative improvement, and rejection. During deployment, these states guide the learned crossover, mutation, and fusion operations without requiring updates to the model parameters or additional objective function evaluations to maintain these states. By employing ranking, normalized fitness, and the relative geometric characteristics of the population, the method ensures that the structure of the core operator parameters remains independent of the decision space dimensionality. The primary experiments involve training on three 10-dimensional synthetic proxy functions. Separate task-specific training is evaluated on planar-arm optimization and MNIST neuroevolution. Deployment-time component/state interventions and a relative-geometry ablation examine the mechanisms.

Critical strengths

This is a practical mechanism for feedback-based adaptation that relies on fixed, pre-learned parameters. By leveraging the parent-offspring comparison information already generated during the selection process, the mechanism eliminates the need for additional objective evaluations when updating its state. Integrating this into the processes of population mixing, mutation, and fusion enables a pre-trained optimizer to respond dynamically within short search horizons. Furthermore, the core model architecture is relatively compact, and its parameter dimensionality is independent of the dimensionality of the decision space. These characteristics are particularly valuable in scenarios where objective evaluations are computationally expensive and online parameter training is either unfeasible or undesirable.

In the main experiments, this method demonstrates strong performance under a clearly defined, unified evaluation budget. The paper also acknowledges the limitations inherent in applied studies limited to only 5 runs and the results from asymmetric budget experiments presented in the appendix.

Critical weaknesses

The claimed novelty needs clearer qualification, and the incremental contribution over closely related methods is insufficiently established. The abstract’s broad characterization of prior offline learned optimizers as lacking retained search feedback needs qualification. For instance, LES already leverages historical evolution-path statistics to regulate the update process while keeping model parameters fixed. Furthermore, B2Opt has already achieved the learning and integration of evolutionary operators, while POM utilizes relative fitness information to facilitate dimension-independent population mixing. Consequently, the contribution of GIFT-Evo should be more accurately described. To substantiate its incremental value, more rigorous controlled ablation studies are required, along with a comparison against POM—a method that is cited in the text but omitted from the experimental section.

Questions for authors

What are the metrics and implementation details for F20? Do the negative values in Table 1 represent the raw objective values or the error relative to the optimal value?

In the MNIST experiments, a CoordinateAdapter is additionally incorporated. Can its core contribution be isolated? How much MNIST performance remains without that adapter?

Appendix A.2 appropriately limits the translation-equivariance statement to crossover. Given that the fusion coefficients need not sum to one, how sensitive is the complete optimizer to consistent translations of the objective, initialization, and search domain?

The component and state ablations intervene on the trained Full checkpoint without retraining. As shown in Table 11, incorporating search states solely during the fusion phase is sufficient to achieve performance levels on various functions that are nearly on par with the Full version. Given consistent training conditions and evaluation budgets, how does it compare to two alternative variants trained from scratch: (i) one that excludes search state conditions entirely, and (ii) one that includes search state conditions only during the fusion phase? These comparisons help quantify the benefits derived from historical feedback.

1 comment
AuthorsOpenReview ↗

For thanks

Thank you for your careful review and constructive suggestions. Your comments have helped us identify where the paper needs clearer positioning and stronger experimental support.

In a future revision, we will clarify the contribution relative to prior work, strengthen the comparisons with POM, and add independently retrained state ablations. We will also validate the F20 implementation and reporting metrics, isolate the contribution of the MNIST CoordinateAdapter, and examine the translation sensitivity of the complete optimizer. We will revise the method and conclusions as needed based on these findings.

We appreciate your feedback, which will help us improve the rigor and clarity of our work.

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

Related papers

Discussion (0)

Sign in to comment.