acceptodds
Under review as a conference paper at ICLR 2027

From Prediction to Ranking in Offline Model-Based Optimization

Abstract

Offline model-based optimization (MBO) aims to discover high-performing designs using only a fixed dataset of past evaluations. Most existing methods learn a surrogate by pointwise regression, implicitly assuming that accurate value prediction is sufficient for reliable optimization. However, once the optimizer reaches a high-quality design, successful optimization depends on whether the surrogate ranks it above the remaining candidates, rather than on predicting their objective values accurately. To formalize this idea, we separate optimizer reachability from surrogate-based selection and define an optimization-oriented ranking risk between -optimal and non--optimal candidates. We show that this ranking risk directly controls final selection error, whereas pointwise regression reaches the same optimization target only indirectly through regional prediction errors and the objective-margin distribution. The resulting bound further identifies the mismatch between training pairs and optimizer-induced target pairs as a key source of error, and relates this mismatch to the geometric separation of high-quality candidates from the offline data support. Guided by this analysis, we propose distribution-aware ranking (DAR), which concentrates training on high-quality-versus-remaining pairs to better match the optimization-relevant comparison distribution. Experiments on five Design-Bench tasks validate the predicted extrapolation difficulty and show that DAR achieves the best average rank among the evaluated methods.

Primary area:
optimization
Keywords:
Offline Model-Based Optimization, Pairwise Ranking, Distribution Shift

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dkG8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies offline MBO through ranking errors between near-optimal and inferior candidates. Its analysis separates candidate reachability from final selection and connects selection error to training-pair mismatch. It proposes DAR, which emphasizes comparisons between high-quality observations and the remaining data, with some comparisons within the high-quality subset. Experiments include Branin and five Design-Bench tasks.

Critical strengths
  1. I find the focus on which comparisons matter useful. Separating reachability from selection makes the scope of the ranking argument clearer and provides a reasonable motivation for concentrating training on promising designs.

  2. I like the simplicity of DAR. It uses existing labels and a straightforward sampling change, and achieves competitive results across continuous and discrete tasks with shared hyperparameters and ten-seed reporting.

Critical weaknesses
  1. I am not convinced that the experiments isolate the proposed contribution beyond RaM. The comparison changes both the ranking loss and pair sampling, while the ablations never compare DAR against global sampling with the same hinge loss. The theory also does not establish that DAR reduces the target-distribution mismatch relative to global sampling. Moreover, best-of-128 ground-truth evaluation does not directly test the surrogate-based selection studied by the theory.

  2. I find the baseline comparison incomplete, particularly for the closest prior work. Table 1 uses RaM-RankCosine but omits RaM-ListNet, whose published means exceed DAR on three of the five tasks. Recent methods such as SPADE [1] and established baselines such as COMs [2] and BDI [3] are also absent. Given the small improvements over the included RaM variant, I need stronger comparisons to judge the practical advance.

[1] https://arxiv.org/abs/2605.11246 [2] https://arxiv.org/abs/2107.06882 [3] https://arxiv.org/abs/2209.07507

Questions for authors
  1. Could you compare global pairing and DAR pairing while keeping the hinge loss, initialization, training budget, and search settings identical? This would help isolate the benefit of the proposed sampling strategy.

  2. Why is RaM-ListNet omitted? Does DAR retain an advantage against it and SPADE under matched evaluation settings? Please also identify which Table 1 results were reproduced versus inherited.

  3. Could you report the true score of the surrogate-selected candidate, alongside the best-of-128 score? This would directly test whether the proposed training improves the selection behavior analyzed in the paper.

5 comments
AuthorsOpenReview ↗

Response by the authors (1/3)

Thank you for your comments. Below please find our responses.

W1-1 & Q1 Global pairing vs. DAR pairing under the same hinge loss

W1 (excerpt): I am not convinced that the experiments isolate the proposed contribution beyond RaM. The comparison changes both the ranking loss and pair sampling, while the ablations never compare DAR against global sampling with the same hinge loss.

Q1: Could you compare global pairing and DAR pairing while keeping the hinge loss, initialization, training budget, and search settings identical? This would help isolate the benefit of the proposed sampling strategy.

Thank you for suggesting this ablation. We run a controlled comparison keeping the pairwise hinge loss (β=0.4\beta=0.4), model initialization, training pair budget, and search settings identical, changing only global uniform pairing vs. DAR pairing. Results are reported in the table below (and visualized in https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r3_global_vs_dar.png).

Task (same β=0.4 hinge)Global Inv. ↓DAR Inv. ↓Global Top-20% ρ ↑DAR Top-20% ρ ↑Global Best-of-128 ↑DAR Best-of-128 ↑
Ant3.82%0.80%0.4290.7590.9150.937
D'Kitty2.87%0.65%0.3810.7560.9430.957
Superconductor3.49%2.54%0.5740.6710.5080.502
TF-Bind-819.41%2.15%0.1580.4700.9820.985
TF-Bind-102.65%0.66%0.6010.7890.6410.672
Mean6.45%1.36%0.4290.6890.7980.811

Replacing global pairing with DAR reduces top-versus-rest inversion error (6.45% to 1.36% on average) and improves within-top Spearman rank correlation (0.429 to 0.689) on every task, improving average Best-of-128 score (0.798 to 0.811) under identical loss and budget. We have added this comparison to Section 5.3 and Appendix I.3, separate from the hyperparameter ablations. Thank you very much.

W1-2 Theoretical condition for reducing target-distribution mismatch

W1 (excerpt): The theory also does not establish that DAR reduces the target-distribution mismatch relative to global sampling.

Thank you for pointing this out. For a fixed target distribution QεQ_\varepsilon with finite first moments under the same pair metric, let

e=W1(Qε,Q^τ),d=W1(Q^τ,Qglobal),a=λW1(Q^τ,Qintra).e = W_1(Q_\varepsilon, \widehat Q_\tau), \qquad d = W_1(\widehat Q_\tau, Q_{\mathrm{global}}), \qquad a = \lambda W_1(\widehat Q_\tau, Q_{\mathrm{intra}}).

By the triangle inequality on the mixture QDAR=(1−λ)Q^τ+λQintraQ_{\mathrm{DAR}} = (1-\lambda)\widehat Q_\tau + \lambda Q_{\mathrm{intra}}, we have W1(Qε,QDAR)≤e+aW_1(Q_\varepsilon, Q_{\mathrm{DAR}}) \leq e + a, while the reverse triangle inequality gives W1(Qε,Qglobal)≥d−eW_1(Q_\varepsilon, Q_{\mathrm{global}}) \geq d - e. Thus, d>2e+ad > 2e + a is sufficient for W1(Qε,QDAR)<W1(Qε,Qglobal)W_1(Q_\varepsilon, Q_{\mathrm{DAR}}) < W_1(Q_\varepsilon, Q_{\mathrm{global}}): whenever the empirical cross-region proxy Q^τ\widehat Q_\tau is closer to QεQ_\varepsilon than half its distance to global pairing (up to the λ=0.1\lambda=0.1 within-top term aa), DAR strictly reduces Wasserstein mismatch. We have added this derivation to Appendix E. Thank you very much.

AuthorsOpenReview ↗

Response by the authors (2/3)

W2 & Q2 Comparisons with RaM-ListNet, SPADE, COMs, BDI, and DynAMO, and result provenance

W2: I find the baseline comparison incomplete, particularly for the closest prior work. Table 1 uses RaM-RankCosine but omits RaM-ListNet, whose published means exceed DAR on three of the five tasks. Recent methods such as SPADE [1] and established baselines such as COMs [2] and BDI [3] are also absent. ...

Q2: Why is RaM-ListNet omitted? Does DAR retain an advantage against it and SPADE under matched evaluation settings? Please also identify which Table 1 results were reproduced versus inherited.

Thank you for pointing these out. Because DAR's training objective is fundamentally a pairwise ranking loss whereas ListNet is a listwise loss, our initial submission selected RaM-RankCosine for a fair comparison under pairwise supervision. Following your and Reviewer miZM's suggestion, we expand our evaluation to 29 methods across 12 tasks (5 Design-Bench, 4 GTOPX trajectory, and 3 ViennaRNA tasks), adding RaM-ListNet [4], BDI [3], COMs [2], DynAMO [5], and SPADE [1] under matched 128-candidate evaluation (when official repositories such as SPADE do not release per-task hyperparameter configurations, tuning is performed strictly on an 80/20 held-out split of each task's offline dataset without oracle queries). Across all 29 methods and 12 tasks (complete 29 × 12 per-task table in https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md and average-rank chart in https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r2_12task_avg_rank.png), DAR+ ranks 1st overall (6.58) and DAR ranks 2nd, outperforming other methods while avoiding ListNet's listwise training overhead, as shown below.

MethodAntD'KittySuperconTF-Bind-8TF-Bind-1012-Task Avg. Rank (of 29) ↓
DAR+ (Ours)*0.916±0.0370.964±0.0100.481±0.0100.991±0.0000.660±0.0286.58
DAR (Ours)*0.941±0.0130.959±0.0060.506±0.0080.988±0.0080.686±0.0568.00
ROOT*0.958±0.0120.967±0.0010.451±0.0320.977±0.0150.652±0.0208.58
RaM-RankCosine*0.915±0.0260.949±0.0140.522±0.0200.976±0.0230.663±0.0478.83
BDI*0.958±0.0000.941±0.0000.506±0.0090.973±0.0000.629±0.0368.83
RaM-ListNet*0.928±0.0130.962±0.0120.516±0.0300.963±0.0180.654±0.02011.08
DynAMO*0.874±0.0220.940±0.0040.517±0.0210.882±0.0930.557±0.03213.83
COMs*0.947±0.0260.952±0.0020.516±0.0170.744±0.1300.625±0.02318.71
SPADE*0.689±0.1320.849±0.0380.339±0.0430.482±0.0860.487±0.05926.50

Regarding result provenance across the 29-method, 12-task benchmark (https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md), 25 of the 29 methods (including DAR+, DAR, ROOT, RaM-RankCosine, RaM-ListNet, BDI, DynAMO, COMs, SPADE, GA, GA-min, GA-mean, Match-OPT, ICT, DEMO, CMA-ES, CbAS, REINFORCE, BONET, RoMA, DDOM, MINs, BO-qEI, GABO, and FGM):

  • discrete TF-Bind-8/10 for GTG, TF-Bind-10 for Tri-Mentoring, and all GTOPX and RNA-Binding tasks across all 29 methods are executed end-to-end by us (*)
  • only the remaining Design-Bench entries marked with † (PGS and IOM on Design-Bench, GTG on Ant/D'Kitty/Superconductor, and Tri-Mentoring on Ant/D'Kitty/Superconductor/TF-Bind-8) are adopted directly from RaM [4];
  • because training those remaining baselines across all tasks and seeds is computationally resource-intensive and time-consuming, we could not complete every seed within our limited compute budget during the rebuttal period; we ran 1–2 seeds and observed results closely matching RaM's reported scores, so we directly adopted RaM's Design-Bench numbers for those † entries for now, and we will continue the remaining runs and update them in a subsequent revision.

We have updated Section 5.2 and marked these result sources in the caption of Table 5 (Appendix I.2) in the revised paper. Thank you very much.

AuthorsOpenReview ↗

Response by the authors (3/3)

W1-3 & Q3 True score of the surrogate-selected candidate

W1 (excerpt): Moreover, best-of-128 ground-truth evaluation does not directly test the surrogate-based selection studied by the theory.

Q3: Could you report the true score of the surrogate-selected candidate, alongside the best-of-128 score? This would directly test whether the proposed training improves the selection behavior analyzed in the paper.

Thank you for this question. Alongside Best-of-128 (Sbest=max⁡x∈C128f(x)S_{\mathrm{best}} = \max_{x \in C_{128}} f(x)), we evaluate candidate-batch median quality and the true score of the single candidate chosen by the frozen surrogate (Ssel=f(arg⁡max⁡x∈C128h(x))S_{\mathrm{sel}} = f(\arg\max_{x \in C_{128}} h(x))), including DAR+ with early-epoch synthetic calibration (applying the synthetic pair mixture during the first 50 training epochs and real DAR pairs during the remaining 50 epochs). Results across the five Design-Bench tasks are reported in the table below (and plotted in https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r4_surrogate_selection.png).

Evaluation MetricMethodAntD'KittySuperconTF-Bind-8TF-Bind-10Avg. Rank ↓
Median-of-128MSE0.559±0.0200.882±0.0020.357±0.0120.510±0.0230.468±0.0082.60
Median-of-128RaM-ListNet0.581±0.0160.885±0.0030.348±0.0090.557±0.0260.465±0.0072.20
Median-of-128DAR (Ours)0.562±0.0210.887±0.0020.374±0.0160.585±0.0280.472±0.0061.20
Surrogate Top-1 (SselS_{\mathrm{sel}})MSE0.586±0.2370.873±0.0240.352±0.0390.817±0.1500.469±0.0772.40
Surrogate Top-1 (SselS_{\mathrm{sel}})RaM-ListNet0.732±0.2140.877±0.0140.341±0.0540.862±0.0900.458±0.0432.00
Surrogate Top-1 (SselS_{\mathrm{sel}})DAR (Ours)0.596±0.2230.881±0.0320.367±0.0330.723±0.1890.489±0.0501.60
Surrogate Top-1 (SselS_{\mathrm{sel}})DAR+ (Ours)0.626±0.1840.925±0.0380.376±0.0310.905±0.0530.520±0.0431.20

DAR achieves the best average rank on both candidate-batch median quality and single-candidate selection among the pure forward surrogates, and DAR+ further improves SselS_{\mathrm{sel}} on every task. We have added these metrics alongside Best-of-128 in Section 5.3 and Appendix I.3. Thank you very much.

References

[1] Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization. https://arxiv.org/abs/2605.11246

[2] Conservative Objective Models for Effective Offline Model-Based Optimization. ICML 2021. https://arxiv.org/abs/2107.06882

[3] Bidirectional Learning for Offline Infinite-Width Model-Based Optimization. NeurIPS 2022. https://arxiv.org/abs/2209.07507

[4] Offline Model-Based Optimization by Learning to Rank. ICLR 2025.

[5] Michael S. Yao, James C. Gee, and Osbert Bastani. Diversity By Design: Leveraging Distribution Matching for Offline Model-Based Optimization. ICML 2025.

[6] ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge. NeurIPS 2025.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

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

This paper proposed a novel method to address the offline black-box optimization problem in which the target is finding high-performing candidates by using only a historical dataset. Conditioned on reaching a near-optimal candidate and stated sampling assumptions, this paper first derived a bound on the probability of selecting an inferior design using ranking error. The authors then introduced a new approach, distribution-aware ranking (DAR), which emphasizes comparisons between high-value samples and the remaining ones. Empirically, DAR showed good performance on Branin and Design Bench tasks.

Critical strengths
  • This paper proposed a bound on the probability of finding a low-performing sample using ranking loss, conditioned on reaching a near-optimal candidate. This bound also exposed the mismatch between training pairs and optimizer-induced target pairs, and linked this mismatch to the geometric data coverage.
  • The proposed method attained good performance on Branin and 5 tasks of the Design-Bench benchmark.
Critical weaknesses
  • The two subsets Sτ+S_{\tau}^+ and Sτ−S_{\tau}^- may not empirically coincide with the true regions in Eq. (2). In particular, the offline dataset only contains low-value samples; thus, using the threshold τ\tau to partition the offline dataset into Sτ+S_{\tau}^+ and Sτ−S_{\tau}^- will cause Sτ+S_{\tau}^+ not to include high-performing samples as expected in the true region.
  • The experiments are not sufficient to prove the efficiency of the proposed methods over existing baselines. Several previous baselines are not compared, which makes the DAR's current average ranking incorrect.
Questions for authors
  1. Regarding the first point in Weaknesses, I am wondering whether using synthetic data can help boost the method's performance. Specifically, instead of using the offline dataset, the authors can try to synthesize more samples to expand the coverage of the data. Data can be synthesized by using another baseline to find candidates and assigning the baseline's prediction as pseudo-labels, or by using synthetic data generation from ROOT [1].

  2. Several previous baselines are not compared in the experiment, including CMA-ES [2], ICT [3], BDI [4], BONET [5], DEMO [6], and GABO [7]. The average ranking of DAR might be lower when including these baselines' results. For example, CMA-ES showed significant performance on the Ant task, which is consistently shown in previous papers [1], [8]. In addition, the authors seem to report the results of RaM [8] with RankCosine loss, which is slightly worse than RaM with ListNet loss on the continuous tasks like Ant and D'Kitty. Therefore, if RaM with ListNet loss is also included, the average rank of DAR will further degrade.

  3. Recent papers also conducted experiments on additional datasets and benchmarks, for instance, RNA-Binding [1][9] or SOO-Bench [10][11]. Comparing DAR against other baselines on these datasets will further prove DAR's efficiency.

[1] Dao, Cuong, Phi Le Nguyen, Truong Thao Nguyen, and Nghia Hoang. "Root: Rethinking offline optimization as distributional translation via probabilistic bridge." Advances in Neural Information Processing Systems 38 (2026): 98783-98817.

[2] Hansen, Nikolaus. "The CMA evolution strategy: a comparing review." Towards a new evolutionary computation: Advances in the estimation of distribution algorithms (2006): 75-102.

[3] Yuan, Ye, Can Sam Chen, Zixuan Liu, Willie Neiswanger, and Xue Steve Liu. "Importance-aware co-teaching for offline model-based optimization." Advances in Neural Information Processing Systems 36 (2023): 55718-55733.

[4] Can Chen, Yingxue Zhang, Jie Fu, Xue Liu, and Mark Coates. Bidirectional learning for offline infinite-width model-based optimization. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022.

[5] Mashkaria, Satvik Mehul, Siddarth Krishnamoorthy, and Aditya Grover. "Generative pretraining for black-box optimization." In International Conference on Machine Learning, pp. 24173-24197. PMLR, 2023.

[6] Ye Yuan, Youyuan Zhang, Can Chen, Haolun Wu, Melody Zixuan Li, Jianmo Li, James J. Clark, and Xue Liu. Design editing for offline model-based optimization. Transactions on Machine Learning Research, 2025.

[7] Michael S Yao, Yimeng Zeng, Hamsa Bastani, Jacob R. Gardner, James Gee, and Osbert Bastani. Generative adversarial model-based optimization via source critic regularization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.

[8] Rong-Xi Tan, Ke Xue, Shen-Huan Lyu, Haopu Shang, Yao Wang, Yaoyuan Wang, Fu Sheng, and Chao Qian. Offline model-based optimization by learning to rank. In The Thirteenth International Conference on Learning Representations, 2025

[9] Ronny Lorenz, Stephan H. Bernhart, Christoph Höner Zu Siederdissen, Hanna Tafer, Christoph Flamm, Peter F. Stadler, and Ivo L. Hofacker. ViennaRNA Package 2.0. Algorithms for Molecular Biology, 6(1):26, November 2011.

[10] Qian, Hong, Yiyi Zhu, Xiang Shu, Shuo Liu, Yaolin Wen, Xin An, Huakang Lu, Aimin Zhou, Ke Tang, and Yang Yu. "SOO-bench: Benchmarks for evaluating the stability of offline black-box optimization." In International Conference on Learning Representations, vol. 2025, pp. 81720-81784. 2025.

[11] Tan, Rong-Xi, Ming Chen, Ke Xue, Yao Wang, Yaoyuan Wang, Sheng Fu, and Chao Qian. "Towards universal offline black-box optimization via learning language model embeddings." arXiv preprint arXiv:2506.07109 (2025).

6 comments
AuthorsOpenReview ↗

Response by the authors (1/3)

Thank you for your comments. Below please find our responses.

W1 & Q1 Low-value offline subsets and synthetic data motivated by ROOT (DAR+)

W1: The two subsets Sτ+S^+_\tau and Sτ−S^-_\tau may not empirically coincide with the true regions in Eq. (2). In particular, the offline dataset only contains low-value samples; thus, using the threshold τ\tau to partition the offline dataset into Sτ+S^+_\tau and Sτ−S^-_\tau will cause Sτ+S^+_\tau not to include high-performing samples as expected in the true region.

Q1: Regarding the first point in Weaknesses, I am wondering whether using synthetic data can help boost the method's performance. ... or by using synthetic data generation from ROOT [1].

Thank you for raising this point and suggesting the connection to ROOT [1]. When the offline dataset SS contains only low-value designs, Sτ+S_\tau^+ serves as an observable proxy rather than coinciding with Xε\mathcal{X}_\varepsilon. Extending our ranking framework explains when synthetic trajectory augmentation motivated by ROOT [1] bridges this gap and guides the design of DAR+:

  1. Coupled reachability, target comparisons, and training comparisons. Let GG denote candidate generation and hh the ranking surrogate. Making (G,h)(G,h) explicit in Eqs. (5)–(9) under Theorem 3 and Corollary 5(a) gives:

Pr⁡(FG,h)≤δε,G,h+KG,hEε,G,h(h)≤δε,G,h+KG,h[E^tr(h)+ΓN+LW1(Qε,G,h,Qtr)].\Pr(F_{G,h}) \leq \delta_{\varepsilon,G,h} + K_{G,h}\mathcal{E}_{\varepsilon,G,h}(h) \leq \delta_{\varepsilon,G,h} + K_{G,h}\Bigl[\widehat{\mathcal{E}}_{\mathrm{tr}}(h) + \Gamma_N + L W_1(Q_{\varepsilon,G,h}, Q_{\mathrm{tr}})\Bigr].

Here FG,hF_{G,h} is failure to return an ε\varepsilon-optimal design, δε,G,h\delta_{\varepsilon,G,h} bounds non-reachability, KG,hK_{G,h} bounds non-ε\varepsilon-optimal candidates, and Qε,G,hQ_{\varepsilon,G,h} is the optimizer-induced target comparison distribution. Selection risk therefore couples reachability, optimizer-induced target comparisons, and training comparisons: when offline data contain only low-value designs, pairs drawn solely from SS can leave a large mismatch W1(Qε,G,h,Qtr)W_1(Q_{\varepsilon,G,h}, Q_{\mathrm{tr}}) relative to candidates reached during search.

  1. Coverage expansion vs. synthetic preference flips. By Corollary 6, for a fixed target positive marginal ρε+\rho_\varepsilon^+ and Dout(M)=Ex∼ρε+dist⁡(x,M)D_{\mathrm{out}}(M)=\mathbb{E}_{x\sim\rho_\varepsilon^+}\operatorname{dist}(x,M), nested supports satisfy M0⊆M1⇒Dout(M1)≤Dout(M0)M_0 \subseteq M_1 \Rightarrow D_{\mathrm{out}}(M_1) \leq D_{\mathrm{out}}(M_0). Meanwhile, on augmented pairs, let Q~aug\widetilde Q_{\mathrm{aug}} assign synthetic preferences and Qaug⋆Q_{\mathrm{aug}}^\star assign true preferences; if they disagree with probability πflip\pi_{\mathrm{flip}}, any loss in [0,B][0, B] satisfies sup⁡h∣EQ~augℓ~h−EQaug⋆ℓ~h∣≤Bπflip\sup_h \lvert\mathbb{E}_{\widetilde Q_{\mathrm{aug}}}\widetilde\ell_h - \mathbb{E}_{Q_{\mathrm{aug}}^\star}\widetilde\ell_h\rvert \leq B\pi_{\mathrm{flip}}, yielding:

Eε,G,h(h)≤E^aug(h)+ΓN+LW1(Qε,G,h,Qaug⋆)+Bπflip.\mathcal{E}_{\varepsilon,G,h}(h) \leq \widehat{\mathcal{E}}_{\mathrm{aug}}(h) + \Gamma_N + L W_1(Q_{\varepsilon,G,h}, Q_{\mathrm{aug}}^\star) + B\pi_{\mathrm{flip}}.

Expanding positive support monotonically reduces off-support extrapolation distance DoutD_{\mathrm{out}}, whereas synthetic preference errors add an explicit flip penalty BπflipB\pi_{\mathrm{flip}}—so synthetic augmentation must expand decision-relevant coverage (W1W_1) while keeping preference flips (πflip\pi_{\mathrm{flip}}) small.

  1. Concrete DAR+ design motivated by ROOT. Guided by this bound and motivated by ROOT [1], DAR+ fits an offline RBF-kernel Gaussian process purely on the offline dataset SS with zero extra oracle calls and forms MGP=8M_{\mathrm{GP}}=8 perturbed GP posterior-mean functions f~(r)r=18\\{\tilde f^{(r)}\\}_{r=1}^8. From anchors in Sτ+S_\tau^+, short gradient ascent and gradient descent rollouts are generated separately under each f~(r)\tilde f^{(r)}. Pairs are formed strictly within each individual f~(r)\tilde f^{(r)} (never across r≠r′r \neq r', avoiding uncalibrated cross-function offsets that inflate πflip\pi_{\mathrm{flip}}): (a) intra-function ascent-vs.-descent endpoint pairs ordered by f~(r)\tilde f^{(r)}, and (b) intra-function neighboring-step trajectory pairs along a single rollout ordered by f~(r)\tilde f^{(r)}. Each update mixes real DAR pairs (τ=0.2,λ=0.1\tau=0.2, \lambda=0.1) with intra-function ascent-vs.-descent endpoint pairs and intra-function neighboring-step trajectory pairs.

In our evaluation in Table R1 (visual rank chart in https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r2_12task_avg_rank.png and complete per-task scores in https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md), DAR+ ranks 1st overall (6.58), improving over both DAR (8.00) and other baselines. We have added these derivations in Section 4 and Appendix F, and full DAR+ specifications in Section 5.2 and Appendix G of the revised paper. Thank you very much.

AuthorsOpenReview ↗

Response by the authors (2/3)

W2 & Q2 & Q3 Expanded comparison across 29 methods and 12 tasks

W2: The experiments are not sufficient to prove the efficiency of the proposed methods over existing baselines. Several previous baselines are not compared, which makes the DAR's current average ranking incorrect.

Q2: Several previous baselines are not compared in the experiment, including CMA-ES [2], ICT [3], BDI [4], BONET [5], DEMO [6], and GABO [7]. ... In addition, the authors seem to report the results of RaM [8] with RankCosine loss, which is slightly worse than RaM with ListNet loss on the continuous tasks like Ant and D'Kitty. ...

Q3: Recent papers also conducted experiments on additional datasets and benchmarks, for instance, RNA-Binding [1][9] or SOO-Bench [10][11]. Comparing DAR against other baselines on these datasets will further prove DAR's efficiency.

Thank you for suggesting these baselines and benchmark tasks. Following your suggestions, we evaluate 29 methods across 12 tasks—spanning morphology control, superconductor materials, DNA/RNA binding sequence design [1, 9], and spacecraft trajectory optimization from SOO-Bench / GTOPX [10, 11]—adding CMA-ES [2], ICT [3], BDI [4], BONET [5], DEMO [6], GABO [7], RaM-ListNet [8], COMs [12], DynAMO [13], and SPADE [14]. Overall average ranks across the 12 tasks are summarized in Table R1 below.

Table R1: Overall average rank across 12 tasks and 29 methods (lower is better; * denotes methods executed by us across all 12 tasks, † denotes Design-Bench entries adopted from RaM [8], and *† indicates mixed Design-Bench sources).

RankMethodAvg. Rank ↓RankMethodAvg. Rank ↓
1DAR+ (Ours)*6.5816GTG*†14.67
2DAR (Ours)*8.0017PGS†15.58
3ROOT*8.5818CbAS*16.08
4RaM-RankCosine*8.8319REINFORCE*17.50
5BDI*8.8320BONET*17.67
6GA*9.5021RoMA*17.67
7GA-min*9.7522DDOM*18.42
8RaM-ListNet*11.0823COMs*18.71
9GA-mean*11.5824MINs*18.83
10Tri-Mentoring*†13.0825BO-qEI*19.96
11Match-OPT*13.1726IOM†21.50
12DynAMO*13.8327GABO*22.08
13ICT*14.2528FGM*24.08
14DEMO*14.2529SPADE*26.50
15CMA-ES*14.42–––

Visual 29-method average rank chart: https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r2_12task_avg_rank.png; complete 29 × 12 per-task mean ± std table and task details: https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md.

Across the 12-task benchmark, DAR+ ranks 1st overall (6.58) and DAR ranks 2nd, outperforming other methods. While optimizers such as CMA-ES perform strongly on individual tasks like Ant, DAR and DAR+ maintain more consistent rankings across different methods.

Note on RNA-Binding and ROOT: Because the official ROOT repository [1] did not release RNA-Binding task code, data preprocessing, or RNA-specific hyperparameters, we construct RNA-A/B/C following BootGen [15] (5,000 length-14 RNA sequences scored by the ViennaRNA binding oracle [9]) and run ROOT using its released discrete TF-Bind-8 hyperparameter configuration adapted to sequence dimension 42 under the matched 128-candidate evaluation, which accounts for the difference between our reproduced ROOT numbers and those reported in [1]. Similarly, on SOO-Bench / GTOPX [10], because existing offline MBO baselines assume an unconstrained domain without box bounds, bounded search variables are mapped bijectively to an unbounded logit domain following [11] so all 29 methods optimize under a unified unconstrained interface (full per-task results and protocols in https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md).

We have updated Section 5.2 and Appendix I.2 with the 29-method, 12-task comparison and task protocols. Thank you very much.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

AuthorsOpenReview ↗

Response by the authors (3/3)

References

[1] ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge. NeurIPS 2025.

[2] The CMA Evolution Strategy: A Comparing Review. Towards a New Evolutionary Computation 2006.

[3] Importance-Aware Co-Teaching for Offline Model-Based Optimization. NeurIPS 2023.

[4] Bidirectional Learning for Offline Infinite-Width Model-Based Optimization. NeurIPS 2022.

[5] Generative Pretraining for Black-Box Optimization. ICML 2023.

[6] Design Editing for Offline Model-Based Optimization. TMLR 2025.

[7] Generative Adversarial Model-Based Optimization via Source Critic Regularization. NeurIPS 2024.

[8] Offline Model-Based Optimization by Learning to Rank. ICLR 2025.

[9] ViennaRNA Package 2.0. Algorithms for Molecular Biology 2011.

[10] SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization. ICLR 2025.

[11] Ji Deng, Jun Gao, Zhao Li, and Ji Zhang. EfficientRefiner: An Efficient Refinement Method over Black-Box Optimization in Macro Placement. https://openreview.net/forum?id=KBNvtUfjZL

[12] Conservative Objective Models for Effective Offline Model-Based Optimization. ICML 2021.

[13] Michael S. Yao, James C. Gee, and Osbert Bastani. Diversity By Design: Leveraging Distribution Matching for Offline Model-Based Optimization. ICML 2025.

[14] Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization. ICML 2026.

[15] Bootstrapped Training of Score-Conditioned Generator for Offline Design of Biological Sequences. NeurIPS 2023.

Reviewer_6TYuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that offline model based optimization (MBO) should focus on ranking promising designs correctly rather than predicting objective values accurately. A major strength is its framework separating candidate reachability from surrogate selection, which cleanly isolates sources of optimization failure. Furthermore, the theoretical analysis connects cross region ranking risk to final selection error, identifying training-pair distribution mismatch and geometric data coverage as central drivers of ranking reliability.

Critical strengths

The paper is clearly written and easy to follow. The authors provide clear definitions of the key concepts and introduce the problem formulation in a structured manner, making the main ideas and theoretical analysis relatively easy to understand.

Critical weaknesses

The method may favor known high-quality regions and miss better solutions elsewhere.

Questions for authors

I have some doubts about the motivation. Why don't we consider other regions with high uncertainty? Could this motivation cause the optimization process to become trapped in a local optimum?

How to identify this true "ε-optimal region" if we don't know the optimum value?

How to justify the relationship between its theoretical $\epsilon$-optimal region and its empirically selected high-quality subset?

The demo example uses Branin as the objective function. Did the author try other objective functions? Branin has three optima with the same value. Identifying one of them could be considered as success optimization. How about using another example with different optimum values?

5 comments
AuthorsOpenReview ↗

Response by the authors (1/2)

Thank you for your comments. Below please find our responses.

W1 & Q1 Exploration of high-uncertainty regions and local optima

W1: The method may favor known high-quality regions and miss better solutions elsewhere.

Q1: I have some doubts about the motivation. Why don't we consider other regions with high uncertainty? Could this motivation cause the optimization process to become trapped in a local optimum?

Thank you for this question. In offline MBO, exploration and selection play different roles than in online BO. Because the true objective cannot be queried during optimization, steering gradient search into high-uncertainty or out-of-distribution regions exposes the optimizer to surrogate overestimation, a central failure mode in offline MBO studied in prior works such as COMs (Trabucco et al., 2021) [1] and GABO (Yao et al., 2024) [2]. Instead, DAR retains the observed top subset (τ=0.2\tau=0.2) to focus supervision on decision-relevant comparisons while initializing parallel gradient search from 128 starting points in Sτ+S_\tau^+.

Equation (3) formalizes this trade-off by separating failure to reach an ε\varepsilon-optimal candidate from failure to select one once reached. When high-value regions are weakly covered in the offline dataset, DAR can be combined with offline coverage expansion (such as offline GP synthetic-trajectory augmentation in DAR+, which ranks 1st across 29 methods and 12 tasks in https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md) rather than unguided exploration. Moreover, our unequal-optima diagnostic in Q4 and the GTOPX trajectory tasks show that DAR is less prone to being trapped in inferior local optima than baseline methods. We have clarified the roles of exploration, coverage, and selection in Sections 2 and 4 of the revised paper. Thank you very much.

Q2 & Q3 True ε-optimal region vs. empirical top-τ subset

Q2: How to identify this true "ε-optimal region" if we don't know the optimum value?

Q3: How to justify the relationship between its theoretical ϵ\epsilon-optimal region and its empirically selected high-quality subset?

Thank you for pointing this out. Algorithmically, DAR constructs Sτ+S_\tau^+ directly from the empirical (1−τ)(1-\tau)-quantile of offline labels and never requires knowing f⋆f^\star or ε\varepsilon. Here Xε\mathcal{X}_\varepsilon is the theoretical target defining selection performance, whereas Sτ+S_\tau^+ is the observable training proxy.

Applying the triangle inequality to their comparison distributions (writing Q^τ\widehat Q_\tau for empirical top-versus-rest pairs) connects them directly:

W1(Qε,QDAR)≤W1(Qε,Q^τ)+λW1(Q^τ,Qintra).W_1(Q_\varepsilon, Q_{\mathrm{DAR}}) \leq W_1(Q_\varepsilon, \widehat Q_\tau) + \lambda W_1(\widehat Q_\tau, Q_{\mathrm{intra}}).

The first term measures how well observable cross-region comparisons represent those needed for final selection, and the second reflects the λ=0.1\lambda=0.1 within-top supervision. Corollaries 5–6 further relate this transfer to distributional and geometric coverage DoutD_{\mathrm{out}}. Thus, relative quality within a low-value dataset supplies directional supervision when its comparison structure transfers to higher-value candidates encountered during search (and our sensitivity study confirms stable performance across intermediate τ∈[0.15,0.25]\tau \in [0.15, 0.25]). We have made this connection explicit in Sections 4 and 5 of the revised paper. Thank you very much.

AuthorsOpenReview ↗

Response by the authors (2/2)

Q4 Objective function with unequal optimum values

Q4: The demo example uses Branin as the objective function. Did the author try other objective functions? Branin has three optima with the same value. Identifying one of them could be considered as success optimization. How about using another example with different optimum values?

Thank you for this suggestion. Following the shifted-data protocol in Section 5.1 (retaining only the worst 60% of offline samples), we add a 2D Tilted Himmelblau diagnostic with four unequal local maxima and a unique global optimum. Results are reported in the table below.

Method (Tilted Himmelblau)Best-of-128 ↑Radius-Mean Inversion ↓Min Dist. to Global Peak ↓
MSE0.526.48%1.045
RaM-RankCosine0.823.74%0.989
DAR (Ours)1.862.93%0.869

DAR outperforms both RaM-RankCosine and MSE. As shown in https://anonymous.4open.science/r/DAR-iclr27/figures/fig_r1_tilted_himmelblau.png, DAR concentrates high surrogate scores around the unique global peak, maintains the lowest cross-region inversion error across all tested extrapolation radii, and approaches the global optimum most closely (with the same ordering holding under larger search step sizes). Furthermore, our expanded 29-method, 12-task benchmark (https://anonymous.4open.science/r/DAR-iclr27/table_r1_full_results.md) includes four GTOPX spacecraft trajectory tasks, showing that DAR remains less prone to inferior local optima beyond 2D.

We have added this unequal-optima diagnostic alongside Branin in Section 5.1 and Appendix I.1 of the revised paper. Thank you very much.

References

[1] Trabucco et al. Conservative Objective Models for Effective Offline Model-Based Optimization. ICML 2021.

[2] Yao et al. Generative Adversarial Model-Based Optimization via Source Critic Regularization. NeurIPS 2024.


We hope our response has addressed your concerns. If we have missed anything, please let us know.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 6TYu,

Thank you very much for taking the time to carefully review our work and provide your constructive comments. Your feedback has been invaluable in helping us improve our work.

In response to your comments, we have included the following content:

  • We clarified why OOD exploration triggers surrogate overestimation in offline MBO, formalized reachability vs. selection in Eq. (3), and showed that offline coverage expansion (DAR+) avoids inferior local optima and ranks #1 across 29 methods and 12 tasks (Sections 2 and 4; Response (1/2)).
  • We confirmed Sτ+S_\tau^+ never requires knowing f⋆f^\star or ε\varepsilon, derived a W1W_1 bound linking QεQ_\varepsilon to the observable proxy Q^τ\hat{Q}_\tau, and verified stability over τ∈[0.15,0.25]\tau \in [0.15, 0.25] (Sections 4 and 5; Response (1/2)).
  • We added the unequal-optima 2D Tilted Himmelblau diagnostic (DAR: 1.86 Best-of-128 vs. 0.82 RaM and 0.52 MSE, with 2.93% inversion error) alongside four higher-dimensional GTOPX trajectory tasks (Section 5.1, Appendix I.1; Response (2/2)).
  • Besides, as suggested by Reviewers dkG8 and miZM, we isolated DAR vs. global pairing under identical hinge loss (Appendix E), expanded our benchmark to 29 methods across 12 tasks (Section 5.2, Appendix I.2), and added surrogate-selected candidate scores (S_sel) and batch medians alongside Best-of-128 (Section 5.3, Appendix I.3).

May we kindly ask if our response has addressed the issues you raised? If you still have any further questions or concerns, we would be more than happy to continue the discussion.

Thank you again for your time and effort.

Best regards,

Authors of Submission 60641

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

Dear Reviewers and Area Chair,

Thank you for your reviews; red main-text updates (moving the full benchmark into Table 1) and new experiments appear in our revised manuscript and anonymous repository.

Theoretical mismatch analysis and synthetic preference extension (dkG8, miZM, 6TYu). We link target regions to empirical top quantiles, prove when DAR reduces mismatch over global pairing, and extend synthetic-preference bounds to offline Gaussian-process rollouts motivating oracle-free DAR+ (Section 4, Appendices E–G).

Expanded baselines and out-of-distribution task suites (dkG8, miZM). Against added ranking, generative, and trajectory baselines (RaM-ListNet, SPADE, BDI, COMs, ICT, BONET, DEMO, GABO, DynAMO, CMA-ES) on ViennaRNA binding and GTOPX trajectory tasks without retuning, DAR+ and DAR rank first and second overall (Section 5.2, Table 1, Appendix I.2).

Controlled pairing ablations and evaluation protocols (dkG8, 6TYu). We isolate boundary-focused from global pairing under matched hinge loss and budget, test unequal optima on Tilted Himmelblau, and verify median-batch and single-candidate selection (Sections 5.1, 5.3, Appendices I.1, I.3).

Other revisions (dkG8, miZM). Alongside main-text Table 1, we document baseline provenance, held-out offline tuning, and task-level variance trade-offs.

Thank you for the constructive discussion.

Best regards,

Authors of Submission 60641

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.