acceptodds
Under review as a conference paper at ICLR 2027

DIAL: Aligning Local Gradients and Global Transport for Few-Shot Offline Optimization

Abstract

Few-shot offline optimization aims to discover high-performing designs (e.g., materials) from sparse labeled data via two main approaches: surrogate and generative methods. The former learns local improvement directions but might not extrapolate well out of the observed data regime. The latter captures global mass-transport structure between low- and high-performing design regimes but might induce generative paths inconsistent with local improvement directions. Despite operating at different scales, we show that these approaches are coupled. Integrating the differential changes in surrogate prediction along a mass-transport path must recover the oracle value difference between its endpoints. Therefore, violations of this relation provide a cooperative correction signal between the surrogate gradient field and generative transport, allowing each model to expose and correct the other's generalization errors and improve overall performance. Yet, in few-shot tasks, this coupling cannot be learned well from the target task alone because sparse and small labeled data provide little access to the differential information required along such paths. To address this challenge, we introduce a **Differential-to-Integral ALignment** (DIAL) meta optimization framework that learns this coupling across a population of similar functions whose behaviors are consistent with the observed target data. The learned coupling can then be transferred to the target task to enrich its few-shot data with shared improvement structures. Our experiments across eight benchmark tasks demonstrate consistent improvements over SOTA methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
offline optimization, generative models, score matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xrTrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DIAL, which couples local surrogate gradients with global generative transport for few-shot offline optimization. It learns differential alignment on GP-based synthetic functions and adapts to the target task through integral alignment along endpoint-conditioned paths between observed designs. The paper reports competitive performance on eight tasks from Design-Bench and SOO-Bench.

Critical strengths
  1. Coupling local surrogate gradients with global transport is an interesting approach to addressing limited supervision. The methodology clearly describes synthetic-task training, target-task adaptation, and the role of alignment in each stage. (§3)
  2. Removing alignment reduces mean performance on the evaluated tasks, supporting the empirical contribution of the proposed training objectives. Comparisons with ROOT under the same GP construction also show additional improvements. (Tables 3, 11)
Critical weaknesses

The paper presents mutual correction of generalization errors as a central explanation, but direct evidence for this mechanism is insufficient. The ablations show that alignment contributes to final optimization performance, but do not establish whether each model’s errors decrease or whether such reductions arise from complementary interactions between the models. A gap therefore remains between the observed performance gains and the proposed error-correction mechanism. (Abstract, §1, Table 3)

Questions for authors
  1. Can the authors analyze each model’s errors before and after joint alignment to directly substantiate the claim that surrogate gradients and generative transport correct each other’s generalization errors?
  2. Can comparisons between updating only one model and jointly updating both models clarify how complementary interactions contribute to error reduction and improved optimization performance?
7 comments
AuthorsOpenReview ↗

Response to Reviewer xrTr

We would like to thank the reviewer for the detailed and constructive feedback. The reviewer’s concern on additional ablation of the error-correction mechanism is addressed below.

Q1 - surrogate gradients and generative transport correct each other’s generalization errors

Given the review comments in the weakness section, our understanding is that the reviewer (1) agrees that the ablation (as reported in Table 3) has demonstrated the contribution of alignment to the optimization performance of the combined test-time sampling method but (2) wants us to further establish that the surrogate and generative models indeed correct each other’s generalization.

We address the additional request in (2) as follows:

First, we note that in addition to Table 3, Appendix C.8 provides results comparing DIAL against several alternative coupling strategies, demonstrating its effectiveness over simpler coupling mechanisms (see details in the appendix).

Second, we evaluate each model’s individual optimization performance with and without alignment training. In the former case, each model was trained in isolation. In the latter case, both models were trained to collaboratively minimize the alignment loss but were evaluated separately, without relying on their combination during test-time optimization.

The results are reported below.

AntDKittyTFBind8TFBind10
Surrogate w/o alignment0.902 ± 0.0340.935 ± 0.0200.828 ± 0.0770.602 ± 0.037
Surrogate w/ alignment0.935 ± 0.0260.937 ± 0.0070.918 ± 0.0480.644 ± 0.020
Generative w/o alignment0.917 ± 0.0440.945 ± 0.0110.941 ± 0.0210.674 ± 0.130
Generative w alignment0.946 ± 0.0130.949 ± 0.0150.956 ± 0.0340.688 ± 0.127

We can observe that the individual test-time optimization performance of both the surrogate and generative models consistently improves across all four Design-Bench tasks when they are trained together using the proposed alignment loss. Importantly, each model is evaluated independently, without relying on the other model during test-time optimization. Therefore, these improvements cannot be attributed merely to their combined inference-time behavior. Instead, they demonstrate that alignment improves each model’s ability to identify high-performing designs beyond the observed data, providing empirical evidence that their interactions during training improve their respective generalization behaviors for the optimization task.

Hence, these findings provide strong support for our proposed mechanism of cooperative error correction between surrogate gradients and generative transport.

Q2 -- comparisons between updating only one model and jointly updating both models

Following the reviewer’s suggestion, we conducted additional ablations comparing updates to only the surrogate model, only the generative flow, or both, under the proposed alignment framework. In each case, the model not being updated is kept frozen during alignment training.

Update strategyAntDKittyTFBind8TFBind10
flow model only0.917 ± 0.0150.935 ± 0.0150.948 ± 0.0270.662 ± 0.043
surrogate model only0.927 ± 0.0300.947 ± 0.0190.945 ± 0.0210.626 ± 0.046
DIAL (both)0.951 ± 0.0130.958 ± 0.0050.961 ± 0.0310.690 ± 0.084

We can observe that DIAL consistently achieves the best performance across all tasks, outperforming both single-model update variants. Importantly, jointly updating both models yields better optimization performance than updating either model alone under the same alignment loss. These results highlight the benefits of complementary interactions between the surrogate model and generative flow during alignment training, further supporting our proposed cooperative error-correction mechanism.

We hope that our responses have sufficiently addressed all your concerns. We would sincerely appreciate your consideration of a higher score if you find these concerns resolved. If you have any further questions or require additional clarification, please do not hesitate to let us know. We would be happy to continue the discussion and provide further details.

AuthorsOpenReview ↗

Follow-up on Reviewer Concerns

Dear Reviewer xrTr,

As we are now halfway through the rebuttal period, we would like to follow up on your remaining concerns. We appreciate your questions regarding the performance of individual models before and after joint alignment, as well as the effects of updating each model independently.

As detailed in our responses, we have provided additional experimental results that directly address these questions and offer further empirical evidence supporting the effectiveness of our alignment framework. In particular, these results demonstrate the benefits of joint alignment over independently updating either model, further reinforcing our original findings.

We would appreciate hearing whether these additional results have addressed your concerns. If any questions remain, we would be happy to continue the discussion and provide further clarification.

Thank you again for your time and consideration.

Best regards,

The Authors

Thank you for the response and additional experiments. The Design-Bench comparisons between updating only one model and jointly updating both models support the benefit of joint adaptation for optimization performance and directly address my second question. Moreover, the improved standalone performance of each model after alignment shows that the benefits cannot be attributed solely to combining the two models at inference time. These additional results substantially alleviate my main concerns.

Regarding my first question, however, the additional experiments evaluate each model’s optimization performance rather than directly comparing its generalization errors before and after alignment. The Introduction states that “the complementary information captured by one view can constrain the other, identifying and correcting inconsistencies between the local and global views,” while the Abstract describes this as “allowing each model to expose and correct the other’s generalization errors.” Although mutual constraint and reduced inconsistency are consistent with the alignment design, improved consistency should be distinguished from a reduction in actual generalization errors. The additional results support the benefits of complementary joint training, but provide indirect evidence for the specific mechanism of exposing and correcting each other’s errors.

I therefore suggest either adding an error analysis appropriate to each model’s role before and after alignment, if feasible, or revising the wording, particularly in the Abstract, to emphasize the consistency encouraged by alignment and the empirically observed performance improvements. My remaining concern is about matching the strength of the mechanistic explanation to the available evidence, rather than the practical effectiveness of the method itself. Thank you.

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

This paper tackles Few-Shot Offline Optimization: how to combine a surrogate model’s local gradient information with a generative model’s global generative capability when labeled data is very limited. The authors propose DIAL, which jointly optimizes both models by enforcing a differential-to-integral consistency relationship between local gradients and global generative paths.

The method has two stages. First, Gaussian processes generate synthetic optimization tasks; a surrogate gradient field and a generative flow are meta-learned, with a differential alignment loss lining up local performance changes with generative transport directions. Second, the meta-trained models are adapted to the target task: Doob’s h-transform builds endpoint-conditioned paths, and an integral alignment loss matches cumulative predicted performance changes along those paths to the observed endpoint differences. Experiments on eight tasks from Design-Bench and SOO-Bench show competitive results, and ablation studies look at how much the main components contribute.

Critical strengths
  1. The problem formulation makes sense, and the motivation is clear. I also think the core idea of combining local surrogate gradients with global generative transport is interesting.

  2. The paper is generally well presented, and the figures are clear and easy to follow. For example, Figure 1 gives an intuitive overview of the proposed method and helps readers understand the overall framework.

  3. The experiments cover eight tasks from two benchmarks, with comparisons against many existing baselines. The results are generally good, and the ablation studies provide additional evidence for the effectiveness of the method.

Critical weaknesses
  1. The core idea is interesting, but I have concerns about its theoretical justification. Since the path integral of a conservative gradient field depends only on the endpoints, it is unclear how the integral alignment loss can effectively optimize the generative transport when the endpoints are fixed. This seems important, as the alignment mechanism is the paper’s main contribution.

  2. The method integrates several existing techniques—Gaussian processes, MAML, flow matching, and Doob’s h-transform. While the overall framework is interesting, the technical novelty of the individual components appears somewhat limited. Moreover, some ablation gains are relatively small, leaving it unclear whether the added complexity is fully justified.

Questions for authors
  1. In Eq. (10), if the endpoints are fixed, how does the integral alignment loss still provide an effective optimization signal for the generative transport? Please clarify how the gradients with respect to θ are obtained.

  2. Is the surrogate gradient field directly parameterized as a vector field, or computed as the gradient of a scalar surrogate function? If it is directly parameterized, how is the consistency relationship in Eq. (1) guaranteed?

4 comments
AuthorsOpenReview ↗

Response to Reviewer VDjo – part 1/2 : Weakness 1 & Question 1

We would like to thank the reviewer for the detailed feedback and address the reviewer's questions as follows:

W1 & Q1 - how the path integral alignment loss optimize the generative transport when the endpoints are fixed

We thank the reviewer for raising this point. This has already been discussed in Remark 1 (immediately following Eq. 1, page 2). Essentially, we emphasize that our alignment loss (via path integral) does depend on the generative transport parameters θ\theta. As already noted in Remark 1 (following Eq. 1): Although the integral of a conservative gradient field depends only on its endpoints, the endpoints reached by the learned generative transport uθ(x(t),t)u_\theta(x(t), t) are themselves determined by the generative transport parameter θ\theta.

Furthermore, as described in Section 3.2.2, we sample terminal endpoints from the transport-induced distribution restricted to the offline dataset:

$ \pi_{\mathcal D}(\mathbf{x}(1)\mid\mathbf{x}(0)) \propto p_\theta^{0\rightarrow 1}(\mathbf{x}(1)\mid\mathbf{x}(0)) \mathbb{I}[\mathbf{x}(1)\in\mathcal D], $

where pθ0→1(x(1)∣x(0))p_\theta^{0\rightarrow 1}(x(1) | x(0)) denotes the endpoint distribution induced by the posterior SDE flow, and I(.)I(.) restricts the sampled designs to the offline dataset.

As such, rather than uniformly sampling offline designs, the endpoint distribution is weighted by the learned generative transport, introducing an explicit dependence on θ\theta. Consequently, the distribution of endpoint pairs used in the alignment loss also depends explicitly on θ\theta, establishing the dependence of the expected alignment loss on the generative transport parameters. To empirically demonstrate that our alignment loss is sensitive to the generative transport parameters, and that updating these parameters meaningfully contributes to the resulting optimization performance, we conduct additional ablation experiments. We freeze the generative transport parameters during Differential Alignment (DA), Integral Alignment (IA), or both, while retaining updates to the surrogate model.

Update strategyAntDKittyTFBind8TFBind10
DA: surrogate only; IA: both0.905 ± 0.0370.947 ± 0.0190.951 ± 0.0240.658 ± 0.096
DA: both; IA: surrogate only0.932 ± 0.0370.948 ± 0.0100.945 ± 0.0290.671 ± 0.039
DA & IA: surrogate only0.927 ± 0.0300.947 ± 0.0190.945 ± 0.0210.626 ± 0.046
DIAL (DA & IA: both)0.951 ± 0.0130.958 ± 0.0050.961 ± 0.0310.690 ± 0.084

We can observe that disabling generative transport updates during either alignment stage consistently degrades optimization performance across all four tasks, with further degradation when transport updates are disabled in both stages. These results further strengthen the evidence that the generative transport does get optimized via minimizing the alignment loss (hence, the dependence).

AuthorsOpenReview ↗

Response to Reviewer VDjo – part 2/2 : Weakness 2 & Question 2

W2 - the technical novelty of the individual components appears somewhat limited

Fundamental Novelty. While we agree that DIAL builds upon established mathematical techniques, as virtually any modern machine learning method does, we would like to emphasize that our fundamental contribution is to broaden the methodological design space of offline optimization through a new mathematical perspective that has not been considered or exploited in prior work. In essence, DIAL provides a new perspective on offline optimization by viewing it as a problem of differential cooperative learning between two interacting players, surrogate gradients and generative transport, where their underlying differential-to-integral consistency provides a principled mechanism for mutual correction and joint optimization.

This lays the groundwork for a broader class of cooperative optimization methods for problem classes including offline black-box optimization, offline reinforcement learning, and AI for scientific discovery. Some example new research questions include:

  1. Can we explore improved mechanisms beyond our alignment loss to characterize the interactions between surrogate gradients and global generative transport and jointly optimize them?

  2. Can this perspective be generalized to multiple interacting predictive and generative models, or extended to other decision-making problems, such as aligning Q-value predictions with generative policies in offline reinforcement learning?

  3. More broadly, can such cooperative learning principles provide new foundations for next-generation AI for Science, enabling more reliable and data-efficient design and discovery in complex scientific domains?

We believe these questions open promising directions and long-lasting research agenda that extend well beyond the specific framework developed and empirically evaluated in this paper.

Technical Novelty. To realize this new perspective, we develop several technical components, including population-level differential alignment, target-task integral alignment, and an endpoint-constrained posterior flow sampling method based on Doob's h-transform. In particular, we introduce a novel Kolmogorov backward consistency regularizer for transition-score estimation, whose derivation and empirical validation are provided in Appendix D. While some individual ablation gains are modest, the consistent contributions of the alignment components (Table 3), together with DIAL's leading mean rank across both benchmark suites (Tables 1–2), demonstrate the practical value of the proposed framework.

Q2 - surrogate gradient field

Our surrogate gradient field sϕ(x)=∇gϕ(x)s_\phi(x)=\nabla g_\phi(x) is parameterized as the gradient of a scalar-valued neural network gϕ(x)g_\phi(x) with respect to xx, rather than being modeled as an independent vector field.

The implementation can also be verified in models/gradient_matching.py of our released code.

This formulation provides better flexibility, as the surrogate is not restricted to gradient matching loss and can readily accommodate alternative surrogate training losses, as demonstrated in Appendix C.4. We will add this implementation detail to the main text to avoid potential confusion.

We hope our responses have sufficiently addressed your remaining concerns. If you have any further questions or require additional clarification, please do not hesitate to let us know. We would be happy to continue the discussion and provide further details.

Thank you to the authors for the response. These answers have addressed my questions. Since I have already provided positive feedback, I will keep my score unchanged.

AuthorsOpenReview ↗

Thank you for positive feedback!

Dear Reviewer VDjo,

Thank you for the positive response. We are glad that our responses have addressed your remaining concerns. We sincerely appreciate your time and thoughtful evaluation of our work.

Best regards,

The Authors

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

This paper proposes Differential-to-Integral Alignment (DIAL), a meta-optimization framework for few-shot offline optimization. The method consists of two stages. First, population-level pre-optimization construct a population of closed-form functions that are consistent with the target function around the observed data region. Second, target-task adaptation transfers and further fine-tunes the pre-optimized coupling by requiring the accumulated differential changes along endpoint-conditioned paths to match the observed performance differences between evaluated designs. Experiments on eight tasks from Design-Bench and SOO-Bench demonstrate strong performance against numerous existing baselines. Additional ablation studies and experiments on endpoint-conditioned posterior flow sampling further support the effectiveness of the proposed framework.

Critical strengths
  1. The paper presents a well-motivated mathematical framework with generally clear derivations. The overall methodology is completely structured and technically sophisticated.

  2. The idea of coupling surrogate gradients with generative transport through differential-to-integral alignment is interesting and provides a novel perspective on few-shot offline optimization.

  3. The empirical evaluation is comprehensive, covering eight optimization tasks and a broad set of representative baselines. Tables 1 and 2 show that DIAL achieves the best mean rank on both benchmarks, while the ablation studies further support the effectiveness of the proposed components.

Critical weaknesses
  1. The proposed framework primarily relies on continuous gradient fields and ODE-based transport. However, the experiments also include discrete optimization tasks such as TFBind8 and TFBind10. The paper provides limited explanation of how these continuous operations are applied to discrete design spaces.

  2. Although the ablation studies demonstrate the benefits of differential and integral alignment, it remains unclear whether the improvements require the proposed alignment objectives or could be achieved using simpler alternatives, such as stronger gradient matching.

  3. The pre-optimization stage relies on synthetic functions constructed using Gaussian processes. However, these functions may not accurately represent the true objective landscape, particularly in regions with limited observations. The robustness of DIAL to mismatches between synthetic and target functions remains insufficiently explored.

Questions for authors
  1. How are the continuous gradient fields and ODE-based transport implemented for discrete tasks such as TFBind8 and TFBind10? Could the authors clarify how generated continuous states are converted into valid discrete designs?

  2. Could the authors compare the proposed approach with simpler alternatives, such as stronger gradient matching, to further demonstrate their additional benefits?

  3. How sensitive is DIAL to the quality of GP-generated synthetic tasks? Could the authors provide additional analysis or experiments to evaluate its robustness when the synthetic functions differ substantially from the target objective?

2 comments
AuthorsOpenReview ↗

Response to Reviewer LjvE - part 1/2 Weakness 1-2 & Question 1-2

We would like to thank the reviewer for the detailed feedback and address the reviewer's questions as follows:

W1 & Q1 - discrete optimization tasks

Our paper follows the standard preprocessing protocol established in [1] and widely adopted in offline optimization, which transforms discrete optimization tasks into continuous optimization problems.

For discrete tasks such as TFBind8 and TFBind10, each DNA sequence is represented by a sequence of categorical variables, each taking one of four possible nucleotide values. Each categorical variable is first converted into a one-hot vector, which is then interpolated with a uniform categorical distribution to obtain a smoothed probability vector p\mathbf{p}. This is subsequently mapped into a continuous logit representation z\mathbf{z}, where z(k)=log⁡p(k)−log⁡p(0)z^{(k)}=\log p^{(k)}-\log p^{(0)} for k=1,…,K−1k=1,\ldots,K-1.

Thus, offline optimization operates in the continuous space of concatenated representations [z1,z2,…,zd][\mathbf{z}_1,\mathbf{z}_2,\ldots,\mathbf{z}_d], where zi\mathbf{z}_i corresponds to the continuous encoding of the (i)-th categorical variable. After optimization, the generated continuous designs are mapped back into discrete sequences by selecting the category with the highest logit at each position (with the reference category assigned a logit of zero).

This preprocessing procedure is consistent with prior work and can be verified in the implementation of [1] at design-bench/design_bench/datasets/discrete_dataset.py, particularly the to_logits and to_integers functions.

W2 & Q2 - the benefits of differential and integral alignment

We would like to emphasize that we have already compared DIAL against most recent gradient matching baselines in offline optimization -- see Tables 1-2. This includes MATCH-OPT [2] and OptBias [3]. For convenience, we quote the relevant results below, which show that DIAL consistently outperforms both methods across all benchmark tasks.

MethodAntD'KittyTFBind8TFBind10
MATCH-OPT0.869 ± 0.0370.934 ± 0.0100.763 ± 0.1090.615 ± 0.037
OptBias0.906 ± 0.0290.924 ± 0.0190.876 ± 0.0500.635 ± 0.034
DIAL0.951 ± 0.0130.958 ± 0.0050.961 ± 0.0310.690 ± 0.084

Alternatively, Appendix C.4 (Table 9) investigates alternative surrogate-training losses by incorporating the corresponding losses from LTR and COMs into DIAL. The results show that our alignment framework consistently improves upon both baselines across all four Design-Bench tasks, further demonstrating that its benefits are not limited to a particular surrogate-training loss. Moreover, Table 3 ablates the individual contributions of Differential Alignment (DA) and Integral Alignment (IA). Removing either component reduces optimization performance.

AuthorsOpenReview ↗

Response to Reviewer LjvE - part 2/2 Weakness 3 & Question 3

W3 & Q3 - How sensitive is DIAL to the quality of GP-generated synthetic tasks

We would like to explain that our GP-generated tasks are not intended to accurately reconstruct the true function landscape. Rather, they represent a population of plausible functions from which our framework learns shared optimization patterns that generalize beyond individual synthetic functions. This helps create a population-level optimization bias that enriches the limited offline data, following the high-level spirit of meta learning.

Importantly, the subsequent target-transfer phase further adapts the learned meta models to the actual offline dataset, mitigating potential discrepancies between synthetic and target functions. As demonstrated from results in Appendix C.5, both phases contribute to the final optimization performance.

To further investigate robustness to variations in the GP task distribution, we vary the RBF kernel lengthscale from its default value of 1.0 to 2.0, 3.0, and 5.0. Since the lengthscale controls the smoothness of GP-generated functions, increasing it can lead to overly smooth approximations that fail to capture local variations in the true function landscape, potentially degrading the quality of the synthetic optimization tasks. As shown below, DIAL consistently outperforms OptBias [3] and ROOT[4], other GP-based meta-training methods, across all settings on the Ant task. Moreover, DIAL exhibits substantially smaller performance degradation as the lengthscale increases, demonstrating its robustness to potential inaccuracies in the GP-generated task distribution.

lengthscale1.02.03.05.0
OptBias0.906 ± 0.0290.870 ± 0.0330.881 ± 0.0180.861 ± 0.024
ROOT0.926 ± 0.0290.900 ± 0.0240.897 ± 0.0290.888 ± 0.024
DIAL0.951 ± 0.0130.948 ± 0.0160.947 ± 0.0130.939 ± 0.023

References :

[1] Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization, ICML 2022

[2] Learning Surrogates for Offline Black-Box Optimization via Gradient Matching, ICML 2024

[3] Black-Box Optimization From Small Offline Datasets via Meta Learning with Synthetic Tasks, AISTATS 2026

[4] ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge, Neurips 2025

We hope our responses have sufficiently addressed your remaining concerns. If you have any further questions or require additional clarification, please do not hesitate to let us know. We would be happy to continue the discussion and provide further details.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear AC and Reviewers,

We sincerely thank you for the detailed/constructive feedback and the thorough review process.

In our rebuttal, we addressed the following key points:

  1. Clarification of implementation details. We have provided additional explanation regarding how offline optimization methods on continuous input space can be applied to discrete tasks following prior practice in [1,2], as requested by Reviewer LjvE. We have also explained how the path integral alignment loss optimizes the generative transport in response to Reviewer VDjo's inquiry. The reviewer finds our answer satisfactory. We note that our implementation is sufficiently detailed in the main text and can be verified through our released code.

  2. Clarification of novelty and contributions. We have also highlighted the novelty of our proposed framework, from conceptual to technical contributions. We also outlined several potential fundamental research directions enabled by our new perspective as discussed in our responses to Reviewer VDjo. We are glad that the reviewer finds our answer satisfactory. We have also emphasized that our experiments in the current manuscript already demonstrate significant improvements over alternative gradient matching approaches (Tables 1-2, Table 9 in Appendix C4) in our responses to Reviewer LjvE.

  3. Further empirical ablations. We have provided additional ablation results to further demonstrate the effectiveness of our method. Specifically, we investigated the sensitivity to the Gaussian process task population in response to Reviewer LjvE. In response to Reviewer xrTr, we provided additional ablations showing the individual performance improvements of each model when their training is coupled within our alignment framework, as well as comparisons between updating only one model and jointly updating both models. We also highlighted several ablation studies already presented in our Appendix, which further strengthen the empirical evidence supporting our approach.

Overall, we believe that these clarifications and additional results have addressed all concerns raised by the reviewers. We thank the reviewers for their insightful questions.

References :

[1] Design-Bench: Benchmarks for Data-Driven Offline Model-Based Optimization, ICML 2022

[2] Conservative Objective Models for Effective Offline Model-Based Optimization, ICML 2021

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.