acceptodds
Under review as a conference paper at ICLR 2027

BiaWDRO: Controlling OOD Error in Offline Model-Based Optimization via Wasserstein Distributionally Robust Optimization

Abstract

Offline model-based optimization (MBO) seeks to discover high-performing designs that maximize a black-box objective function using only pre-collected datasets, with applications such as materials discovery, biological sequence design, and robot morphology optimization. The central challenge in offline MBO is out-of-distribution (OOD) error, where surrogate models yield overconfident predictions that mislead the optimization process toward spurious optima. To mitigate this risk, conservative methods have emerged as a prevalent approach. However, these methods often rely on heuristic techniques. We introduce BiaWDRO, which adapts Wasserstein Distributionally Robust Optimization (DRO) to offline MBO by controlling worst-case predictive bias within Wasserstein ambiguity sets. Applying Lagrangian duality, we derive an objective that mitigates OOD overestimation through tractable adversarial perturbations of training samples. This approach establishes connections between DRO and conservative modeling, providing a principled approach that replaces previous heuristic methods. We further derive an explicit lower bound on empirical fitting error, quantifying the trade-off between worst-case predictive bias and fitting accuracy. Experiments on Design-Bench demonstrate competitive performance across diverse offline MBO tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Offline Model-Based Optimization, Scientific Design Optimization, Distributionally Robust Optimization, Conservative Optimization

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_c8tTOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BiaWDRO, a conservative surrogate model learning framework for Offline Model-Based Optimization (Offline MBO), aimed at mitigating the problem of surrogate models producing overly optimistic predictions on out-of-distribution (OOD) data, which can lead the optimization process toward spurious optima. The core idea of BiaWDRO is to define and control the Worst-case Predictive Bias, which measures the maximum increase in the mean prediction of a surrogate model under distributional shifts constrained by the Wasserstein distance. The authors incorporate this metric as a regularization term into the surrogate model's training objective to suppress unreliable, excessively high predictions. Specifically, by leveraging Wasserstein duality, they transform the challenging distributional optimization problem into an adversarial sample search problem with a transportation cost penalty. Adversarial samples are generated through gradient ascent, while the surrogate model parameters and dual variables are updated alternately to train a more conservative surrogate model. Furthermore, the authors theoretically analyze the trade-off between conservatism and model fitting accuracy and conduct experiments on five Design-Bench tasks. The experimental results demonstrate that BiaWDRO achieves competitive performance on offline design optimization tasks.

Critical strengths
  1. The paper addresses surrogate overestimation on out-of-distribution (OOD) samples in offline model-based optimization. It introduces a Wasserstein DRO-based conservative objective and provides a theoretical analysis of the trade-off between conservatism and prediction accuracy.

  2. The method is evaluated on five Design-Bench tasks covering both continuous and discrete design spaces. BiaWDRO achieves a competitive overall average rank, suggesting its potential effectiveness across different optimization settings.

Critical weaknesses

1.The proposed method has substantial overlap with existing work, particularly Conservative Objective Models (COMs), which already use gradient-based adversarial samples and conservative regularization to mitigate surrogate overestimation. Moreover, Wasserstein duality, transport-cost penalties, and distributionally robust adversarial training have been extensively studied in prior work, such as Sinha et al. (2018). The main contribution appears to be introducing a Wasserstein transport penalty into a COMs-style training framework, with a dual variable controlling adversarial exploration. However, the paper does not clearly establish what fundamentally new capability this formulation provides over existing approaches, making the claimed methodological novelty somewhat limited.

2.More importantly, the current ablation studies do not adequately validate the core technical contribution. Although the paper examines different transport costs and Wasserstein radii, these experiments mainly investigate design choices rather than demonstrate the necessity of the Wasserstein DRO formulation. In particular, comparisons against COMs augmented with simple distance regularization, fixed transport penalties, or adaptive transport penalties are missing. It therefore remains unclear whether the improvements arise from the proposed DRO formulation or simply from regularizing the distance traveled by adversarial samples. Furthermore, the contributions of the adaptive dual variable and the adaptive conservatism weight are not separately evaluated. Without these comparisons, the empirical evidence is insufficient to establish the benefits of the proposed formulation over simpler alternatives.

3.The effectiveness of the proposed worst-case control mechanism remains insufficiently validated. While the theoretical objective considers the worst-case prediction increase within a Wasserstein uncertainty set, the practical algorithm approximates this objective using finite-step local gradient ascent. As the surrogate evolves, new spurious high-prediction regions may emerge, and it is unclear whether the adversarial search can reliably identify and suppress them. Moreover, controlling prediction increases does not necessarily reduce overestimation errors relative to the true objective. Although Figure 1 illustrates the intended suppression of spurious peaks, the experiments mainly report aggregate optimization scores and exploration distances, without directly demonstrating that this mechanism occurs in practice.

Questions for authors
  1. What is the essential technical contribution beyond COMs and existing Wasserstein DRO methods?

Since COMs already uses adversarial samples for conservative surrogate training, and transport-cost-based adversarial optimization has been studied previously, what fundamentally distinguishes the proposed method? In particular, could similar improvements be achieved by simply introducing a distance penalty into COMs?

  1. Can the authors provide ablations that directly isolate the benefits of the Wasserstein DRO formulation?

A particularly important comparison would be between COMs with simple distance regularization, fixed transport penalties, and the proposed adaptive transport penalty. Additionally, can the authors independently ablate the adaptive dual variable and the adaptive conservatism weight? These experiments would help determine whether the gains genuinely come from the proposed formulation rather than simpler regularization or adaptive weighting mechanisms.

  1. Can the authors provide more direct evidence that the method suppresses spurious prediction peaks?

Since the worst-case search relies on finite-step local gradient ascent, how sensitive are the results to the number of ascent steps and the use of multiple initializations? Could the authors also provide a low-dimensional synthetic example showing the ground-truth objective, surrogate prediction landscapes before and after conservative training, and optimization trajectories? Together with a direct evaluation of OOD overestimation errors, this would help verify whether the method genuinely suppresses spurious peaks rather than merely shifting their locations.

5 comments
AuthorsOpenReview ↗

Response to Reviewer c8tT (1/3)

Thank you for your careful review and constructive comments, which we address point by point below.

Q1 Fundamental contribution relative to COMs and existing Wasserstein DRO methods

We propose a distributionally robust conservative learning criterion for offline model-based optimization. In supervised learning, WDRO minimizes the maximum expected prediction loss under distribution shift. When inputs are perturbed while their labels are retained, this objective encourages predictions at the perturbed inputs to remain close to the original labels. However, optimizing a design can change its true objective value, so this label preservation can suppress prediction differences that guide the search for better designs. Our key modeling insight is that the maximum increase in expected surrogate prediction over an ambiguity set of input distributions admits a DRO formulation. We use this quantity as a conservative regularizer alongside data fitting, thereby obtaining a DRO learning objective for offline MBO.

Relative to COMs, BiaWDRO uses a Wasserstein ambiguity set to specify the distribution shifts over which its prediction difference regularizer is maximized. This allows more flexible adjustment of conservatism: the transport cost and radius specify the geometry and extent of the distribution shifts being regularized, while the regularization weight controls the penalty strength.

To evaluate the effectiveness of this DRO-derived training scheme, we compare BiaWDRO with COMs augmented by a distance penalty. Specifically, we add −βc(x~,x)-\beta c(\tilde x,x) to COMs' training adversarial objective, where c(x~,x)=∥x~−x∥22/dc(\tilde x,x)=\|\tilde x-x\|_2^2/d, and compare it with fixed-γ\gamma BiaWDRO. Both methods use nine fixed penalty weights: 0.1, 0.3, 0.5, 1, 1.5, 2, 3, 5, and 10.

The complete results are reported in Table 11 (revised PDF, Appendix C.4.3, p. 20) and reproduced below. Scores are the highest true normalized score among the final 128 candidates, averaged across eight seeds. Mean ranks are computed jointly among the 18 settings.

MethodTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
BiaWDRO, fixed γ=20.992 ± 0.0010.651 ± 0.0130.505 ± 0.0040.952 ± 0.0040.954 ± 0.0015.2
BiaWDRO, fixed γ=1.50.990 ± 0.0010.645 ± 0.0070.508 ± 0.0040.946 ± 0.0040.953 ± 0.0056.0
BiaWDRO, fixed γ=50.991 ± 0.0000.667 ± 0.0180.495 ± 0.0090.959 ± 0.0060.949 ± 0.0016.2
BiaWDRO, fixed γ=0.30.981 ± 0.0040.634 ± 0.0130.509 ± 0.0040.950 ± 0.0110.956 ± 0.0036.2
BiaWDRO, fixed γ=0.10.976 ± 0.0050.618 ± 0.0150.512 ± 0.0050.946 ± 0.0050.958 ± 0.0037.4
BiaWDRO, fixed γ=10.981 ± 0.0040.659 ± 0.0150.506 ± 0.0040.941 ± 0.0040.956 ± 0.0047.4
BiaWDRO, fixed γ=30.992 ± 0.0010.633 ± 0.0120.501 ± 0.0050.947 ± 0.0060.953 ± 0.0018.4
COMs + distance, β=20.989 ± 0.0030.651 ± 0.0100.506 ± 0.0040.942 ± 0.0060.946 ± 0.0028.2
BiaWDRO, fixed γ=0.50.987 ± 0.0020.629 ± 0.0070.509 ± 0.0040.942 ± 0.0040.951 ± 0.0038.6
COMs + distance, β=0.30.975 ± 0.0050.635 ± 0.0160.506 ± 0.0040.954 ± 0.0090.945 ± 0.0028.8
BiaWDRO, fixed γ=100.991 ± 0.0000.665 ± 0.0120.488 ± 0.0040.937 ± 0.0080.947 ± 0.00110.0
COMs + distance, β=0.50.969 ± 0.0040.675 ± 0.0280.506 ± 0.0040.944 ± 0.0060.943 ± 0.00110.0
COMs + distance, β=100.443 ± 0.0420.475 ± 0.0180.507 ± 0.0030.948 ± 0.0040.947 ± 0.00111.2
COMs + distance, β=1.50.989 ± 0.0020.644 ± 0.0190.506 ± 0.0040.939 ± 0.0050.943 ± 0.00210.6
COMs + distance, β=50.424 ± 0.0390.483 ± 0.0140.506 ± 0.0040.953 ± 0.0080.945 ± 0.00211.4
COMs + distance, β=0.10.971 ± 0.0050.611 ± 0.0130.505 ± 0.0040.944 ± 0.0070.946 ± 0.00112.6
COMs + distance, β=10.985 ± 0.0020.641 ± 0.0080.505 ± 0.0040.936 ± 0.0090.945 ± 0.00112.2
COMs + distance, β=30.687 ± 0.0720.564 ± 0.0200.505 ± 0.0040.951 ± 0.0050.945 ± 0.00112.4

The six settings with the best mean ranks are all fixed-γ\gamma BiaWDRO variants. The best mean rank is 5.2 for BiaWDRO, compared with 8.2 for distance-regularized COMs. These results support the effectiveness of BiaWDRO's DRO-derived training scheme beyond simply adding a fixed distance penalty to COMs.

AuthorsOpenReview ↗

Response to Reviewer c8tT (2/3)

Q2 Separating the effects of transport regularization and adaptive weighting

Thank you for this suggestion. We added COMs with simple distance regularization and independently ablated the transport dual variable γ\gamma and conservatism weight λ\lambda. All settings use eight random seeds. Scores are the highest true normalized score among the final 128 candidates.

(1) Independent ablation of adaptive γ\gamma. We fix γ\gamma while retaining adaptive λ\lambda and all other BiaWDRO settings. Our method achieves the best mean rank, as shown in Table 9 (revised PDF, Appendix C.4.1, p. 19). Although some fixed weights achieve higher scores on individual tasks, adaptive transport weighting achieves better average performance across the evaluated tasks than fixed weights.

MethodTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
Adaptive γ\gamma0.987 ± 0.0070.713 ± 0.0610.513 ± 0.0130.952 ± 0.0120.956 ± 0.0072.4
Fixed γ=20.992 ± 0.0010.651 ± 0.0130.505 ± 0.0040.952 ± 0.0040.954 ± 0.0014.0
Fixed γ=50.991 ± 0.0000.667 ± 0.0180.495 ± 0.0090.959 ± 0.0060.949 ± 0.0014.8
Fixed γ=0.30.981 ± 0.0040.634 ± 0.0130.509 ± 0.0040.950 ± 0.0110.956 ± 0.0035.0
Fixed γ=1.50.990 ± 0.0010.645 ± 0.0070.508 ± 0.0040.946 ± 0.0040.953 ± 0.0055.8
Fixed γ=30.992 ± 0.0010.633 ± 0.0120.501 ± 0.0050.947 ± 0.0060.953 ± 0.0015.8
Fixed γ=0.10.976 ± 0.0050.618 ± 0.0150.512 ± 0.0050.946 ± 0.0050.958 ± 0.0036.0
Fixed γ=10.981 ± 0.0040.659 ± 0.0150.506 ± 0.0040.941 ± 0.0040.956 ± 0.0045.8
Fixed γ=0.50.987 ± 0.0020.629 ± 0.0070.509 ± 0.0040.942 ± 0.0040.951 ± 0.0037.0
Fixed γ=100.991 ± 0.0000.665 ± 0.0120.488 ± 0.0040.937 ± 0.0080.947 ± 0.0017.2
COMs0.941 ± 0.0480.647 ± 0.0280.465 ± 0.1140.933 ± 0.0300.946 ± 0.00710.0

(2) Independent ablation of adaptive λ\lambda. We fix λ∈{0.01,0.1,1}\lambda\in\{0.01,0.1,1\} while retaining adaptive γ\gamma and all other settings. Although some fixed weights achieve higher scores on individual tasks, our method offers a better balance of performance across tasks and achieves the best mean rank. Table 10 (revised PDF, Appendix C.4.2, p. 19) reports the results, with ranks computed within these four settings from task means; ties receive the minimum shared rank.

Conservatism-weight settingTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
Full adaptive method0.987 ± 0.0070.713 ± 0.0610.513 ± 0.0130.952 ± 0.0120.956 ± 0.0071.2
Fixed λ=0.10.700 ± 0.0360.632 ± 0.0110.433 ± 0.0130.946 ± 0.0070.961 ± 0.0052.6
Fixed λ=0.010.980 ± 0.0030.618 ± 0.0090.500 ± 0.0050.945 ± 0.0050.950 ± 0.0012.8
Fixed λ=10.741 ± 0.0310.666 ± 0.0310.299 ± 0.0300.915 ± 0.0100.706 ± 0.0103.4

(3) Simple distance regularization versus BiaWDRO with fixed transport weights. The complete comparison is provided in the table in our response to Q1 above (Table 11, Appendix C.4.3, p. 20). With both transport weights fixed and adaptive λ\lambda retained, BiaWDRO achieves a better mean rank than distance-regularized COMs. Its advantage therefore does not depend entirely on adaptive transport weighting.

Together, these comparisons support the practical value of the DRO-derived training scheme beyond simple distance regularization and separately demonstrate the contributions of the two adaptive updates.

AuthorsOpenReview ↗

Response to Reviewer c8tT (3/3)

Q3 Direct evaluation of overestimation, search sensitivity, and a low-dimensional example

Thank you for your suggestions. We directly evaluate overestimation along optimization trajectories and examine search sensitivity and a low-dimensional example.

(1) Direct evaluation along optimization trajectories. We examine whether reduced overestimation accompanies improvement in true scores during optimization. On D’Kitty, we compare BiaWDRO and COMs over eight seeds, starting from the same 128 designs within each seed. At each step, both models evaluate the pooled designs generated by both methods, enabling prediction-error comparisons on identical inputs. We also evaluate the true objective values of each method’s 128 candidates at every step and report their mean to track progress in actual design quality.

Figure 8 (revised PDF, Appendix C.6.1, p. 23) presents three complementary views.

  • Panel (a) shows the mean true score of each method’s candidates: COMs plateaus after an initial improvement, whereas BiaWDRO continues to improve and reaches a higher final value.
  • Panel (b) shows the difference in mean prediction error, COMs minus BiaWDRO, on identical inputs, which grows during optimization.
  • Panel (c) directly measures positive overestimation on these same inputs, showing that COMs’ overestimation increases more rapidly.

Together, these results show that BiaWDRO achieves lower overestimation alongside continued improvement in true scores, supporting the practical benefit of balancing conservatism and exploration through predictive-bias regularization.

(2) Sensitivity to the number of inner steps. Across five tasks, we vary the number of gradient-ascent steps in the training inner loop, keeping the other training settings and final design-optimization procedure unchanged. Every configuration uses eight random seeds.

The following table reports the mean across seeds of the highest true normalized score among the final 128 candidates for different numbers of inner steps, KK.

The inner-step sensitivity results are reported in Table 12 (revised PDF, Appendix C.5, p. 20) and reproduced below.

TaskK=1K=1K=5K=5K=10K=10K=20K=20K=50K=50
Ant0.951060.952020.950730.95812—
D’Kitty0.946530.955760.962280.95309—
Superconductor0.507430.513220.482080.48125—
TFBind80.660290.960550.992530.986720.95637
TFBind100.524630.655080.668690.713210.64984

Performance is relatively stable under moderate changes in the number of inner steps, and increasing the step count does not consistently improve performance.

(3) Sensitivity to the number of search starts. We compare one, three, and five search starts at the default number of steps. The first start is the source design; additional starts receive independent Gaussian perturbations with standard deviation 0.15\sqrt{0.15}. Each start returns its final iterate, and the inner objective selects among these results. Transport costs are always calculated relative to the original source design.

The multistart sensitivity results are reported in Table 13 (revised PDF, Appendix C.5, p. 20) and reproduced below.

TaskSingle startThree startsFive starts
Ant0.952020.944980.93840
D’Kitty0.955760.961000.96199
Superconductor0.513220.502660.49472
TFBind80.986720.987590.98582
TFBind100.713210.702650.70070

Across all five tasks, mean true scores vary by less than 0.02 across the multistart settings, indicating limited sensitivity to this setting.

(4) Low-dimensional example: visualizing the relationship between the true objective and search directions. We conducted a 2D toy experiment on the Branin function. The surrogate is trained on approximately 4,000 offline samples (sampled from 5,000 points with the top and bottom 10% removed), intentionally leaving the high-value regions as OOD. Figure 9 (revised PDF, Appendix C.6.2, p. 23) plots optimization trajectories of both COMs and BiaWDRO on the true objective contour. As shown, BiaWDRO finds better optima than COMs while avoiding being misled into the low-value OOD regions (bottom 10%).

Reviewer_s4wHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the out-of-distribution (OOD) overestimation problem in offline model-based optimization (MBO), where surrogate models trained on a static dataset produce overconfident predictions on OOD designs and mislead the gradient-based search toward spurious optima. The authors propose BiaWDRO, which adapts Wasserstein distributionally robust optimization (DRO) to this setting. Instead of minimizing the worst-case loss as in standard DRO, BiaWDRO regularizes the worst-case predictive bias, which is the maximum increase in the surrogate's expected prediction over all distributions within a Wasserstein ball centered at the empirical training distribution. Through Lagrangian duality, the intractable optimization over distributions is reduced to a tractable dual objective in which, for each training point, an adversarial perturbation maximizing the prediction minus a transport cost is found by K steps of gradient ascent. The dual variable γ\gamma is updated by projected gradient descent, and model parameters are updated by alternating stochastic gradients. The authors further prove a lower bound on empirical fitting error (Theorem 3.1) showing that excessive bias regularization collapses training predictions to the mean label, quantifying the conservatism–fitting trade-off. Experiments on five Design-Bench tasks report the best mean rank, ablate the choice of transport cost and the radius ρ, and show via paired distance analysis that BiaWDRO explores less aggressively from initial designs than COMs while achieving higher scores.

Critical strengths
  1. A clean and consequential reformulation of conservatism in offline MBO as worst-case predictive-bias control. The paper's central move regularizing the worst-case increase of the surrogate's expected prediction within a Wasserstein ambiguity set is simple but consequential. This is a genuine conceptual contribution that is likely to be cited and reused.

  2. Informative empirical practice around a fragile evaluation protocol. Design-Bench comparisons are notoriously noisy and the authors engage with this honestly rather than hiding behind a headline number: they report the full 22-method table, verify that their mean-rank claim is robust to task inclusion and tie handling. This substantially raises the confidence that the reported gains, while modest in magnitude, reflect a effective improvement.

Critical weaknesses

The "principled, explicit control" framing overstates what the method actually delivers. The derivation from Eq. (4) to Eq. (9) follows Sinha et al. (2017) essentially step for step, and as the authors themselves show in Sec. 3.2, the trained objective is structurally COMs with a transport-cost penalt. It is a useful unification, but incremental. More importantly, the paper's advertised advantage, which is explicit control of the conservatism level through the Wasserstein radius and the bias-regularization weight, is undercut by the actual algorithm: the regularization weight is not fixed but updated adaptively by an ad hoc schedule, the dual variable is learned by projected gradient descent rather than tied to a target ambiguity level, and Theorem 3.1, the one genuinely new theoretical result, quantifies an intuitively obvious collapse behavior for a fixed-regularization-weight objective that the deployed method does not use. The result is a method whose practical behavior is governed by a handful of coupled hyperparameters (e.g. the Wasserstein radius, the regularization-weight schedule, the dual-variable learning rate, the number of adversarial steps, and the adversarial step size), which is the same heuristic-tuning burden the introduction criticizes in prior work, just relocated.

Questions for authors
  1. How sensitive is the mean-rank result of 3.2/22 to the inclusion of individually weak baselines? If the comparison is restricted to the top-5 or top-10 methods by mean rank, does BiaWDRO remain first, and is the margin statistically meaningful under a paired comparison across seeds (e.g., paired bootstrap or Wilcoxon signed-rank on per-seed scores, rather than comparing SE bars)?

  2. The main text presents the objective with a fixed bias-regularization weight, but Sec. 3.1 states that in experiments the weight is updated adaptively according to Eq. (16) in Appendix C.2. What exactly is the update rule, what triggers increases or decreases, and over what range does the weight vary during training?

  3. What is the wall-clock training cost of BiaWDRO relative to COMs, given the per-batch inner maximization with multiple adversarial steps and the additional dual-variable update?

3 comments
AuthorsOpenReview ↗

Response to Reviewer s4wH (1/2)

Thank you for your careful review and constructive comments, which we address point by point below.

Q1 Sensitivity of mean ranks to weaker baselines

Thank you for suggesting this robustness check. We select the top 10 and top 5 methods according to their mean ranks in the full comparison, then recompute within-task ranks and mean ranks within each subset.

The ranking robustness results for the full, top-10, and top-5 comparison sets are reported in Table 15 (revised PDF, Appendix C.7.1, p. 24) and reproduced below.

Comparison setBiaWDRO mean rankOverall position
All 22 methods3.21st
Top 10 methods2.41st
Top 5 methods2.01st

Detailed results for the top 10 methods are shown below:

The detailed top-10 ranks are provided in Table 16 (revised PDF, Appendix C.7.1, p. 24) and reproduced below.

MethodTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
BiaWDRO112262.4
ROOT2410113.6
RaM-ListNet531823.8
Match-OPT577545.6
NEMO928275.6
RaM-RankCosine3841045.8
BDI394696.2
ICT8106436.2
DDOM759787.2
RoMA10539107.4

Results for the top 5 methods are as follows:

The detailed top-5 ranks are provided in Table 17 (revised PDF, Appendix C.7.1, p. 25) and reproduced below.

MethodTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
BiaWDRO112242.0
ROOT245112.6
RaM-ListNet331522.8
Match-OPT353433.6
NEMO524253.6

BiaWDRO ranks first in all three comparison sets, showing that its leading mean rank does not depend on the inclusion of weaker baselines.

Q2 Statistical significance across random seeds

For comparisons across random seeds, we use eight matched seeds and conduct a paired bootstrap analysis against RaM-ListNet, the second-ranked method in the full table. In each bootstrap sample, we resample the same seed indices for both methods and compute the difference in their mean scores.

The paired score differences and bootstrap confidence intervals are reported in Table 18 (revised PDF, Appendix C.7.2, p. 25) and reproduced below.

TaskMean score difference: BiaWDRO − RaM-ListNet95% confidence interval
TFBind8+0.01972[0.00602, 0.03337]
TFBind10+0.01621[−0.06462, 0.09600]
Superconductor−0.00078[−0.02082, 0.01786]
Ant+0.03102[0.00362, 0.06144]
D’Kitty−0.00624[−0.01784, 0.00506]

Compared with RaM-ListNet, the second-ranked method in the full comparison, BiaWDRO improves the mean scores on TFBind8 and Ant by 0.01972 and 0.03102, respectively, with both 95% confidence intervals lying entirely above zero. Its mean score on TFBind10 also improves. These results indicate that BiaWDRO achieves better overall performance than RaM-ListNet.

AuthorsOpenReview ↗

Response to Reviewer s4wH (2/2)

Q3 Update rule, triggering conditions, and observed range of the bias regularization weight

Thank you for highlighting the relationship between the fixed-weight objective and the adaptive update used in our experiments. The weight λ\lambda in the main objective represents the regularization strength used for the current model update. Between training batches, it is adjusted adaptively according to the predictive bias.

Let L^bias,t=1B∑i=1B[fθt(x~i)−fθt(xi)]\widehat L_{\mathrm{bias},t} = \frac{1}{B}\sum_{i=1}^{B} \left[f_{\theta_t}(\tilde x_i)-f_{\theta_t}(x_i)\right], where x~i\tilde x_i is generated by the adversarial inner optimization. The update rule is λt+1=[λt+ηλ(L^bias,t−τ)]+\lambda_{t+1} = \left[ \lambda_t+ \eta_\lambda\left(\widehat L_{\mathrm{bias},t}-\tau\right) \right]_+.

When the predictive bias exceeds τ\tau, the update direction increases λ\lambda, strengthening the bias penalty. When this estimate is below τ\tau, the update direction decreases λ\lambda, relaxing the restriction on data fitting.

We use τ=0.5\tau=0.5 for the continuous tasks and τ=2\tau=2 for TFBind8 and TFBind10. The ranges observed throughout training across eight random seeds are shown below:

The observed ranges of the conservatism weight are reported in Table 5 (revised PDF, Appendix B.1.1, p. 15) and reproduced below.

TaskMinimum λ\lambdaMaximum λ\lambda
TFBind80.0008600.100000
TFBind100.00001710.100000
Superconductor0.0103760.268951
Ant0.0185310.100000
D’Kitty0.0380460.148094

To further evaluate adaptive λ\lambda, we retain adaptive updates of γ\gamma and fix only λ\lambda to 0.01,0.1,10.01,0.1,1, comparing each setting with the full method. The table reports means ± SE over eight random seeds.

The adaptive-versus-fixed conservatism-weight comparison is reported in Table 10 (revised PDF, Appendix C.4.2, p. 19) and reproduced below.

Conservatism-weight settingTFBind8TFBind10SuperconductorAntD’KittyMean rank ↓
Adaptive λ\lambda0.987 ± 0.0070.713 ± 0.0610.513 ± 0.0130.952 ± 0.0120.956 ± 0.0071.2
Fixed λ=0.1\lambda=0.10.700 ± 0.0360.632 ± 0.0110.433 ± 0.0130.946 ± 0.0070.961 ± 0.0052.6
Fixed λ=0.01\lambda=0.010.980 ± 0.0030.618 ± 0.0090.500 ± 0.0050.945 ± 0.0050.950 ± 0.0012.8
Fixed λ=1\lambda=10.741 ± 0.0310.666 ± 0.0310.299 ± 0.0300.915 ± 0.0100.706 ± 0.0103.4

The adaptive method outperforms all three fixed settings on Ant, Superconductor, TFBind8, and TFBind10. Its five-task mean rank is 1.2, compared with 2.6 for the best fixed setting. These results show that adaptive regularization reduces the need for task-specific weight tuning while improving performance across tasks.

Q4 Actual training time compared with COMs

Thank you for asking about computational cost. We compare COMs and BiaWDRO under the same hardware configuration. The mean training time per epoch is shown below:

The per-epoch training times are reported in Table 7 (revised PDF, Appendix B.1.3, p. 15) and reproduced below.

TaskCOMs (s/epoch)BiaWDRO (s/epoch)
Ant21.174.61
D’Kitty23.305.71
Superconductor39.9110.57
TFBind876.9550.75
TFBind10114.7966.26

BiaWDRO has a lower mean training time per epoch than COMs on all five tasks while maintaining competitive optimization performance. COMs uses more inner gradient-ascent steps to generate high-prediction perturbations, increasing the number of surrogate evaluations and input-gradient computations. For BiaWDRO, shorter transport-regularized inner searches already suffice to achieve strong final scores, as supported by the inner-step ablation in Table 12 (revised PDF, Appendix C.5, p. 20), where additional steps do not consistently improve performance. Fewer inner steps reduce computation and shorten training time per epoch.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer s4wH,

Thank you again for your thoughtful review. We hope our responses and revisions have addressed your concerns. We would be happy to provide any further clarification if helpful.

Reviewer_VzZ6OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

I understand BiaWDRO as a conservative surrogate-learning method for offline MBO. It penalizes the largest increase in expected surrogate predictions within a Wasserstein ball, using transport-penalized adversarial samples during training. The paper also analyzes the fitting cost of excessive conservatism and reports competitive results on five Design-Bench tasks.

Critical strengths
  1. I find the Wasserstein formulation a useful way to formalize conservative training. It gives an explicit transport budget and makes the connection to COMs clear.

  2. I appreciate the analysis of when regularization forces training predictions to collapse to a constant. Together with the radius and cost-function ablations, this helps explain the method’s conservatism–fitting trade-off.

Critical weaknesses
  1. I am not convinced that the proposed “predictive bias” adequately captures OOD overestimation. Eq. (4) measures increases in predictions, without reference to the true objective, so it also penalizes genuine improvements under an accurate surrogate. Theorem 3.1 establishes a fitting-cost bound, rather than an OOD-error guarantee. Likewise, Figure 3 shows shorter displacements, not lower prediction errors. This leaves a gap between the paper’s central motivation and what it actually establishes.

  2. I find the evidence for the empirical advantage incomplete. Appendix B.2 combines reproduced and previously reported results, while acknowledging substantially weaker GTG reproduction without explaining the discrepancy. For example, GTG’s Ant score drops from 0.963 in its original paper to 0.876 here. Recent relevant methods such as SPADE [1] are also missing, and the exploration and small-budget analyses only compare against COMs. These limitations make the best-mean-rank result less convincing to me.

[1] https://arxiv.org/abs/2605.11246

Questions for authors
  1. What connects Eq. (4) to actual OOD overestimation? Do you have measurements of surrogate error along optimization trajectories, beyond displacement from the initial designs?

  2. Appendix C.2 states that the implemented λ update does not certify the proposed bias constraint. How closely does the finite-step procedure satisfy the Wasserstein and bias constraints in practice?

  3. What explains the GTG reproduction gap, and does the empirical advantage remain after resolving it and including recent relevant baselines such as SPADE?

4 comments
AuthorsOpenReview ↗

Response to Reviewer VzZ6 (1/2)

Thank you for your careful review and constructive comments, which we address point by point below.

Q1 Connection between predictive bias and actual OOD overestimation, and direct evaluation along optimization trajectories

(1) Connection between predictive bias and actual OOD overestimation.

In Section 3, predictive bias refers to the increase in the surrogate’s expected prediction under distribution shift. Its connection to actual OOD overestimation is that, under the same distribution shift, the mean prediction error moves toward overestimation when the increase in expected prediction exceeds the increase in the expected true objective. Since true OOD objective values are unavailable during training, predictive bias serves as a proxy for overestimation risk rather than a direct measure of actual overestimation.

Regularizing prediction growth can therefore curb unsupported optimism, but may also suppress predictions for genuinely high-performing designs and limit their exploration. The Wasserstein radius ρ\rho and regularization weight λ\lambda allow us to calibrate this conservatism, balancing protection against overestimation with opportunities for further exploration.

Thank you for suggesting that we measure surrogate-model errors along optimization trajectories. This evaluation will provide more direct evidence of our method’s effectiveness.

(2) Direct evaluation along optimization trajectories.

Following the reviewer’s suggestion, we assess this trade-off by examining whether reduced overestimation accompanies improvement in true scores. On D’Kitty, we compare BiaWDRO and COMs over eight seeds, starting from the same 128 designs within each seed. At each step, both models evaluate the pooled designs generated by both methods, enabling prediction-error comparisons on identical inputs. We also evaluate the true objective values of each method’s 128 candidates at every step and report their mean to track progress in actual design quality.

Figure 8 (revised PDF, Appendix C.6.1, p. 23) presents three complementary views.

  • Panel (a) shows the mean true score of each method’s candidates: COMs plateaus after an initial improvement, whereas BiaWDRO continues to improve and reaches a higher final value.
  • Panel (b) shows the difference in mean prediction error, COMs minus BiaWDRO, on identical inputs, which grows during optimization.
  • Panel (c) directly measures positive overestimation on these same inputs, showing that COMs’ overestimation increases more rapidly.

Together, these results show that BiaWDRO achieves lower overestimation alongside continued improvement in true scores, supporting the practical benefit of balancing conservatism and exploration through predictive-bias regularization.

Q2 Satisfaction of the Wasserstein and bias constraints during training

Thank you for suggesting that we quantify constraint satisfaction during training. We record the transport costs and predictive biases of perturbations generated during training and compare them with the target thresholds. The table below summarizes statistics over the final one-tenth of training batches, averaged across eight random seeds.

The training diagnostics are reported in Table 6 (revised PDF, Appendix B.1.2, p. 15) and reproduced below.

TaskTransport budget ρMean transport costBias threshold τMean predictive bias
Ant1.50.51530.50.4324
D’Kitty1.50.25680.50.3886
Superconductor1.50.26520.50.4773
TFBind81.51.50162.02.0082
TFBind101.51.50102.01.9563

For the three continuous tasks, both the mean transport cost and mean predictive bias are below their respective thresholds. For the discrete tasks, the mean transport cost is close to the budget, with relative deviations of approximately 0.11% and 0.07%, and the mean predictive bias is also close to its target.

The mean transport cost and mean predictive bias near the end of training are below or close to the specified thresholds, indicating that the implementation approximately satisfies both constraints.

AuthorsOpenReview ↗

Response to Reviewer VzZ6 (2/2)

Q3 GTG reproduction gap and comparison with SPADE

Thank you for your suggestions. We provide further details on the GTG reproduction gap and add experimental results and comparisons for SPADE.

(1) Our reproduced GTG score is close to those reported in subsequent work.

Our GTG (NeurIPS 2024) [1] experiment uses the official public implementation and achieves an Ant score of 0.876, compared with 0.963 in the original paper. Among subsequent studies, Offline-RaM (ICLR 2025) [2] and ROOT (NeurIPS 2025) [3] report GTG Ant scores of 0.855 and 0.865, respectively, both close to our result. The comparison table in SPADE (ICML 2026) [4] also uses 0.865, citing ROOT as its source.

The public review discussion of Offline-RaM (ICLR 2025) [2] provides directly relevant information about this gap. In their response to Q4, the authors stated that they used the official GTG code and rechecked the task configurations, trajectory-construction parameters, context length, and conditioning parameter, confirming their reproduction results. The reviewer’s question and the authors’ response are cited in [5, 6].

Our GTG result is therefore comparable to those reported in subsequent studies, and the reproduction discrepancy has already been discussed publicly.

(2) Missing task-specific parameters and additional reproduction results for SPADE.

A difficulty in reproducing SPADE (ICML 2026) [4] is that the paper specifies search ranges for the calibration weight λ1\lambda_1 and support-regularization weight λ2\lambda_2, but does not provide the final parameter values used for each task. This prevents us from fully matching the reported experimental settings.

We first used the public implementation’s defaults, λ1=λ2=1\lambda_1=\lambda_2=1, and completed eight-seed experiments on Ant, D’Kitty, Superconductor, TFBind8, and TFBind10. For these reproduction runs, we used the paper’s training and search budgets of 100 epochs, 100 generations of evolutionary search, and 128 final candidates.

The table below compares our SPADE reproduction results with the reported scores; both are mean ± SE over eight seeds for the maximum normalized oracle score among 128 final candidates. Our reproduced scores are lower on all five tasks.

TaskOur SPADE runsSPADE reported
Ant0.655 ± 0.0220.978 ± 0.006
D’Kitty0.841 ± 0.0260.981 ± 0.003
Superconductor0.351 ± 0.0220.546 ± 0.013
TFBind80.501 ± 0.0350.923 ± 0.015
TFBind100.496 ± 0.0160.915 ± 0.010

Because the final task-specific weights were not released, we tried 16 configurations on Superconductor within the published ranges. The configuration selected by offline validation MSE achieved a score of 0.360. Further adjustments to the inference settings improved the best observed score to 0.424, still below SPADE’s reported 0.546. Since the original task-specific settings remain unavailable, these results may not fully reflect SPADE’s performance under the settings used in the paper.

We therefore use SPADE’s reported scores in the expanded comparison of 23 methods across the five tasks. BiaWDRO retains the best mean rank of 4.0, placing first, while SPADE achieves 4.4, placing second.

References

[1] Taeyoung Yun, Sujin Yun, Jaewoo Lee, and Jinkyoo Park. Guided Trajectory Generation with Diffusion Models for Offline Model-based Optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2024.

[2] Rong-Xi Tan, Ke Xue, Shen-Huan Lyu, Haopu Shang, Yao Wang, Yaoyuan Wang, Sheng Fu, and Chao Qian. Offline Model-Based Optimization by Learning to Rank. In International Conference on Learning Representations (ICLR), 2025.

[3] Manh Cuong Dao, The Hung Tran, Phi Le Nguyen, Thao Nguyen Truong, and Trong Nghia Hoang. ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge. In Advances in Neural Information Processing Systems (NeurIPS), 2025.

[4] Yonghan Yang, Ye Yuan, Zipeng Sun, Linfeng Du, Bowei He, Haolun Wu, Can Chen, and Xue Liu. Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization. In International Conference on Machine Learning (ICML), 2026.

[5] Reviewer wgm9. Review of “Offline Model-Based Optimization by Learning to Rank”. ICLR 2025 OpenReview public discussion, 2024.

[6] Rong-Xi Tan et al. Author response to Q4 on “Offline Model-Based Optimization by Learning to Rank”. ICLR 2025 OpenReview public discussion, November 18, 2024.

I apprciate the authors' response. Most of my concerns are addressed. I updated my rating accordingly.

AuthorsOpenReview ↗

Response to Reviewer VzZ6

Thank you for your thoughtful feedback and for taking the time to consider our response. We’re glad the additional analyses helped address most of your concerns, and we appreciate your reassessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.