- 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
- 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)
- 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
- 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?
- 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
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.
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.
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.
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.
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.
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.