acceptodds
Under review as a conference paper at ICLR 2027

In-Context Preference-Conditioned Pareto Set Learning from Limited Offline Data

Abstract

Multi-objective optimization involves diverse preferences over conflicting objectives. Rather than returning a finite set of trade-off solutions, preference-conditioned Pareto set learning seeks to map a specified preference directly to a corresponding Pareto optimal solution. Yet in many real-world scenarios, optimization must proceed offline without further objective evaluations, leaving only a limited static dataset that sparsely covers the decision and objective spaces. Existing methods either optimize separately for each preference or train a problem-specific Pareto solution model, requiring repeated search across preferences or repeated training across problems. We introduce in-context preference-conditioned Pareto set learning (ICPSL), a framework that amortizes both preference-wise search and problem-specific model construction. ICPSL pretrains a foundation-style hypernetwork on a large and diverse set of synthetic prior data. Given the offline observations of an unseen problem, the pretrained hypernetwork learns to generate a problem-specific, preference-conditioned Pareto solution model in a single forward pass. The generated model can then be reused to predict solutions for arbitrary preferences without problem-specific training or preference-wise search. Experiments on synthetic and real-world problems show that ICPSL produces high-quality Pareto solutions while providing substantial speedups in problem-specific solution-model construction.

Primary area:
optimization
Keywords:
Multi-objective Optimization, Pareto Set Learning, In-context Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dKuiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ICPSL, a Transformer hypernetwork pretrained on synthetic GP-sampled multi-objective problems. Given a small offline dataset, it generates the weights of a preference-conditioned MLP in one forward pass. The MLP outputs a latent-conditioned mixture of logit-normal distributions over solutions, trained with a Tchebycheff free-energy (KL-to-Gibbs) objective.

Critical strengths
  • The two-level amortization (across preferences and across problems) is clearly presented.
  • Objective-based pretraining removes the need for Pareto-optimal labels.
Critical weaknesses
  • The empirical claims are not supported, especially on real-world problems. In Tables 5–7, no ICPSL variant achieves the best HV on RE23, RE24, or RE32 under any context distribution (0/9). ParetoFlow or DOMOO is best in every case. Meanwhile, all of ICPSL-MM's 15 wins come from low-dimensional (d≤3d \le 3) synthetic functions.
  • Prior–target mismatch is not analyzed. The prior draws each objective f~j\tilde{f}_j independently, with no control over the degree of conflict between objectives. The paper does not analyze how this prior–target mismatch affects performance, which may explain the RE results.
  • Ablations are missing. No component is ablated: the Tchebycheff signature embedding versus raw objectives yi\mathbf{y}_i, the size JJ of the reference preference bank, the low-rank decoder, the encoder depth, the temperature τ\tau, or the prior design (kernel mixture, ARD length-scale range).
  • Novelty is limited. The framework combines prior-data hypernetwork pretraining (Müller et al., 2025), preference-conditioned Pareto set learning and hypernetworks (Navon et al., 2021; Lin et al., 2022), and synthetic-GP pretraining for multi-objective optimization (Zhang et al., 2026).

References

Xi Lin, Zhiyuan Yang, Xiaoyuan Zhang, and Qingfu Zhang. Pareto set learning for expensive multi-objective optimization. In Advances in Neural Information Processing Systems, volume 35, pp. 19231–19247, 2022.

Andreas C. Müller, Carlo Curino, and Raghu Ramakrishnan. MotherNet: Fast training and inference via hyper-network transformers. In International Conference on Learning Representations, 2025.

Aviv Navon, Aviv Shamsian, Gal Chechik, and Ethan Fetaya. Learning the Pareto front with hypernetworks. In International Conference on Learning Representations, 2021.

Xinyu Zhang, Conor Hassan, Julien Martinelli, Daolang Huang, and Samuel Kaski. In-context multi-objective optimization. In International Conference on Learning Representations, 2026.

Questions for authors

Please see weaknesses.

2 comments

We thank the reviewer for the constructive comments and suggestions. We respond to each point below.

1. Empirical claims on real-world problems.

Thank you for raising this concern. We agree that Tables 5–7 do not establish HV superiority on the RE problems. Our motivation is to construct a reusable preference-conditioned solution model from sparse offline observations while avoiding costly problem-specific training and repeated preference-wise search. We will clarify that the engineering results support a quality–efficiency trade-off rather than consistent HV superiority.

2. Ablation studies.

Thank you for raising this question. We have added ablation studies by pretraining separate models with different context embeddings, reference preference bank sizes, and ARD length-scale ranges. Following the suggestions raised by Reviewers 8s3L and AgHE about the preference-aligned performance metric, we evaluate these variants using preference-conditioned solution quality (PSQ), defined as the average Tchebycheff value of each predicted solution under its corresponding query preference, which better aligns with query-specific solution quality than HV: PSQ=1T∑q=1Tmax⁡j=1,…,mλqj fj!(x^(λq)).\mathrm{PSQ}=\frac{1}{T}\sum_{q=1}^{T}\max_{j=1,\ldots,m}\lambda_{qj}\ f_j\\!\left(\hat{\boldsymbol{x}}(\boldsymbol{\lambda}_q)\right). Intuitively, lower PSQ indicates better solution quality for the queried preferences. The table reports relative changes from the corresponding default setting: Change(\mathrm{Change}(\\%)=\left[\operatorname{GM}_{c}\left(\frac{\overline{\mathrm{PSQ}}_{\mathrm{ablation},c}}{\overline{\mathrm{PSQ}}_{\mathrm{default},c}}\right)-1\right]\times100\\%, where GM denotes the geometric mean, PSQ‾\overline{\mathrm{PSQ}} is the mean over independent trials and cc indexes matched problem–sampler cases. Negative values indicate improvement.

ComponentAblation vs. Default SettingMAP Change (%)CAM Change (%)MM Change (%)Overall Change (%)
Context embeddingRaw objectives vs. Tchebycheff signature+1.34+1.96+2.21+1.84
Reference preference bankJ=256J=256 vs. J=128J=128−4.52−2.23−1.69−2.82
ARD length-scale range[0.02,0.1][0.02,0.1] vs. [0.1,0.5][0.1,0.5]+55.09+23.43+19.24+31.67
ARD length-scale range[0.5,2.5][0.5,2.5] vs. [0.1,0.5][0.1,0.5]−18.77−6.57−3.49−9.86

The results show that the Tchebycheff signature achieves better PSQ than raw objective inputs under all three inference strategies. This indicates that explicitly representing observation quality across reference trade-offs helps the model infer solutions for queried preferences. Increasing the reference bank size improves PSQ across the three strategies, suggesting that denser trade-off information can refine the context representation. The length-scale ablation suggests that emphasizing more gradually varying objective landscapes during pretraining improves transfer to the evaluated benchmarks. With only a small offline context, such regularity may make the underlying objective structure easier to infer. This highlights the importance of the landscape geometry represented during pretraining and suggests scope for improving prior–target matching.

The low-rank decoder used in our paper keeps GPU memory requirements manageable during pretraining. Generating full-rank hidden-layer weights would sharply increase memory usage, making pretraining infeasible on our available hardware. For encoder depth, we adopt the default 12-layer Transformer encoder used in PFN-based methods. Our comparison with PFNs4BO is reported in Section 1 of our response to Reviewer AgHE. The temperature τ\tau controls the balance between low Tchebycheff energy and distributional entropy. We set τ=0.015\tau=0.015 to favor low-energy solutions while retaining entropy regularization against premature concentration. This value is fixed across target problems without task-specific tuning.

3. Prior–target mismatch.

Thank you for raising this concern. Sampling objective functions independently does not imply that their optima are aligned or that the resulting problems lack conflict. The objectives share the same decision space and can favor different solution regions, producing trade-offs during pretraining. Independent draws from the GP prior allow us to generate a large and diverse collection of training problems, supporting generalization to unseen tasks. Restricting the objectives to prescribed conflict patterns would constrain the allowable function combinations and could reduce this diversity. We therefore adopt independent sampling to encourage broad exposure to different objective landscapes and trade-off structures.

To provide additional analysis on prior–target mismatch, we examine transfer under a controlled extreme example involving objectives with substantially different regularity from the GP prior, as described below. This experiment complements the discussion of objective conflict by testing another important aspect of prior mismatch. We construct the Weierstrass–Takagi multi-objective problem (WTMOP), using Weierstrass [1] and Takagi [2] functions as its two objectives. These functions are continuous but extremely non-smooth, deliberately violating the smoothness of the mixed kernels during training. We use the domain [0,0.5] to ensure clear conflict between the two objectives.

Moreover, the functions have a strict Hölder exponent guarantee with controllable parameters. The Hölder exponent α\alpha measures the local roughness of a function [3,4]. A larger α\alpha indicates a smoother function. For Weierstrass, the α=−log⁡alog⁡b\alpha = -\frac{\log a}{\log b}, we set a=0.7a=0.7, b=3b=3 and α=0.325\alpha=0.325. For Takagi, α=1.0\alpha=1.0. The sample paths of a Matern-ν\nu GP have almost-Hölder exponent ν\nu. The table below compares these Hölder exponents, highlighting the smoothness mismatch between the Weierstrass-Takagi functions and our pretraining distribution.

Function / KernelWeierstrassTakagiMatérn-3/2Matérn-5/2RBF
α\alpha0.3251.01.52.5∞\infty

We evaluate the pretrained model in our paper using ten-observation contexts. The table reports PSQ ranks under LHS, Sobol, and uniform sampling.

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow5.005.006.005.330/30/3
PSLR8.008.008.008.000/30/3
PSLO7.007.007.007.000/30/3
DOMOO6.006.004.005.330/30/3
PFNs4BO1.002.001.001.332/33/3
ICPSL-MAP4.004.005.004.330/30/3
ICPSL-CAM2.001.002.001.671/33/3
ICPSL-MM3.003.003.003.000/33/3

The results show that ICPSL retains competitive preference-conditioned solution quality under this pronounced prior mismatch. CAM outperforms ParetoFlow, PSLR, PSLO, and DOMOO under every context distribution, achieving the best PSQ under Sobol sampling and the second-best under LHS and uniform sampling. MM consistently ranks third. This extreme example suggests that the model can transfer useful information learned during GP pretraining to objectives with substantially different regularity.

[1] Hardy, Godefroy Harold. Weierstrass’s non-differentiable function.

[2] Takagi, Teiji. A simple example of the continuous function without derivative.

[3] Evans, Lawrence C. Partial differential equations.

[4] Gilbarg, David, et al. Elliptic partial differential equations of second order.

4. Novelty clarification.

Thank you for raising this concern. The key contribution of ICPSL is learning to construct a reusable preference-conditioned solution model from an unseen problem’s offline observations in a single forward pass. MotherNet generates classifiers from labeled datasets, existing PSL methods train a solution model for each problem, and TAMO learns a policy for sequential optimization without explicitly modeling the preference–solution structure. ICPSL instead learns how offline observations determine an entire preference-conditioned solution model, amortizing both model construction across problems and solution search across preferences. This capability is relevant to many real-world applications where observations are scarce and decision-makers need to explore different preferences. After synthetic pretraining, ICPSL constructs a solution model without problem-specific parameter updates and answers new preference queries without repeated optimization. This reduces the computational cost of deploying solution models across new problem instances and enables rapid exploration of alternative trade-offs. We will clarify these distinctions and highlight the practical value of our contribution.

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

The paper proposes ICPSL for preference-conditioned multi-objective optimization from limited offline data. A hypernetwork pretrained on synthetic problems generates a problem-specific solution model, which predicts solutions for different preferences without further training or search. Experiments on synthetic functions and engineering design tasks evaluate hypervolume and runtime, showing competitive solution quality and faster model construction.

Critical strengths
  • The paper addresses a clear practical problem: constructing a reusable preference-conditioned solution model from limited offline observations. Generating this model in one forward pass substantially reduces problem-specific training time, which is useful when solving many related problems.

  • Pretraining uses synthetic objective evaluations without requiring Pareto-optimal solution labels. The competitive HV results on several synthetic tasks provide some evidence that this approach can learn useful solution models from sparse observations.

Critical weaknesses
  • The paper does not sufficiently establish how the method handles OOD designs, a central challenge in offline optimization. It remains unclear when pretraining on synthetic GP functions transfers reliably to a new target problem, especially outside regions supported by the offline observations. The experiments mainly use small, randomly sampled contexts and provide little analysis of how performance changes with data coverage or mismatch with the training prior. This leaves the method’s generalization ability insufficiently understood.

  • The experiments cover too few baselines and use only HV to evaluate solution quality. In particular, there is no similarly pretrained baseline to separate the benefit of cross-problem pretraining from that of the proposed architecture. HV alone also does not separately assess diversity, front coverage, or satisfaction of the queried preferences. This makes the comparison with DOMOO incomplete, since balancing diversity and convergence is an explicit goal of that method. These gaps limit the evidence for ICPSL’s advantages over existing approaches.

Questions for authors
  • How does performance change when the target problem differs from the synthetic GP training prior, or when the offline data cover only a limited region? Can you provide evidence that the method handles OOD designs reliably under these conditions?

  • How does ICPSL compare with a model that uses the same synthetic pretraining tasks but directly maps the context and preference to a solution? This comparison would help clarify the benefit of generating a problem-specific model through a hypernetwork.

  • The mixture distribution is motivated by potentially discontinuous optimal solution mappings, but CAM and MM average solutions in decision space. What prevents these averages from falling between disconnected high-quality regions and producing poor solutions? Can you demonstrate the benefit of the mixture formulation on a controlled example?

  • Appendix D.4 uses 10,000 true objective evaluations to determine normalization ranges. Do these ranges serve only to compute HV, or do they also normalize the context inputs? In the latter case, how do the results change when normalization uses only the available offline observations?

  • Does the comparison retain DOMOO’s diversity-driven solution selection stage? Please clarify any adaptations. How do the methods compare in front coverage and preference satisfaction alongside HV, under the same output budget?

If the authors can adequately address these concerns, I am open to revising my score.

5 comments

We thank the reviewer for the constructive comments and suggestions. We respond to each point below.

1. New performance and diversity metrics.

Thank you for the helpful suggestions. We agree that HV alone is not suitable for separately assessing preference satisfaction, diversity, and front coverage. We therefore propose preference-conditioned solution quality (PSQ) and introduce two decision-space diversity metrics, Mean Pairwise Distance (MPD) and Mean Nearest-Neighbor Distance (MND), to better assess preference-aligned performance and solution diversity.

Preference-conditioned solution quality (PSQ) is the average Tchebycheff value of each predicted solution under its corresponding query preference: PSQ=1T∑q=1Tmax⁡j=1,…,mλqjfj(xq)\mathrm{PSQ}=\frac{1}{T}\sum_{q=1}^{T}\max_{j=1,\ldots,m}\lambda_{qj}f_j(x_q), where TT is the number of query preferences, xq=x^(λq)x_q=\hat{x}(\lambda_q) is the predicted solution with query preference λq\lambda_q, and fjf_j is the objective function. Lower PSQ indicates better preference-conditioned solution quality. Compared with HV, PSQ better reflects how well each predicted solution aligns with its corresponding query preference.

For diversity, let x1,…,xT∈[0,1]dx_1,\ldots,x_T\in[0,1]^d be the predicted decisions. Mean Pairwise Distance (MPD) measures their overall dispersion: MPD=2T(T−1)∑i<j∥xi−xj∥2d\mathrm{MPD}=\frac{2}{T(T-1)}\sum_{i < j}\frac{\lVert x_i-x_j\rVert_2}{\sqrt d}. Mean Nearest-Neighbor Distance (MND) measures local crowding: MND=1T∑i=1Tmin⁡j≠i∥xi−xj∥2d\mathrm{MND}=\frac{1}{T}\sum_{i=1}^{T}\min_{j\ne i}\frac{\lVert x_i-x_j\rVert_2}{\sqrt d}. Both normalize distances by d\sqrt d, and larger values indicate greater dispersion or less crowding, respectively. They are complementary because solutions concentrated in two distant clusters can have a large MPD but a small MND. These distances characterize decision-space diversity and do not independently establish Pareto-front coverage or overall optimization quality.

The tables report average ranks over 30 problem–sampler cases, with lower ranks indicating better performance.

PSQ (lower values are better).

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow3.502.553.903.324/3016/30
PSLR6.106.006.606.231/302/30
PSLO5.305.605.505.471/306/30
DOMOO5.405.005.505.303/309/30
PFNs4BO4.105.204.304.537/3012/30
ICPSL-MAP4.254.503.454.071/3011/30
ICPSL-CAM3.853.803.553.730/3017/30
ICPSL-MM3.503.353.203.3514/3018/30

MPD (larger values indicate greater overall dispersion).

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow4.304.104.504.303/3012/30
PSLR6.806.006.006.270/302/30
PSLO7.707.907.907.830/300/30
DOMOO2.002.101.801.9714/3026/30
PFNs4BO2.603.603.003.076/3022/30
ICPSL-MAP3.303.103.703.377/3017/30
ICPSL-CAM4.004.004.104.030/308/30
ICPSL-MM5.305.205.005.170/303/30

MND (larger values indicate less local crowding).

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow7.206.506.906.870/301/30
PSLR4.103.904.004.000/3012/30
PSLO5.906.406.606.300/304/30
DOMOO2.202.102.002.107/3028/30
PFNs4BO1.701.801.601.7023/3027/30
ICPSL-MAP6.706.706.706.700/300/30
ICPSL-CAM3.703.703.503.630/3015/30
ICPSL-MM4.504.904.704.700/303/30

The results show that no single method performs best across all three metrics. ICPSL-MM achieves a PSQ average rank close to ParetoFlow while obtaining the most wins and Top-3 cases. DOMOO performs best in MPD, whereas PFNs4BO performs best in MND. Among the ICPSL variants, MAP achieves greater overall dispersion, while CAM shows less local crowding than MAP and MM. MM provides the strongest preference-conditioned solution quality among our variants, but its predictions are less dispersed. These results reveal differences between preference-conditioned quality and decision-space diversity that HV alone does not separate.

2. Pretrained baseline.

Thank you for the helpful suggestion. We implement a pretrained baseline using PFNs4BO [1] to separate the benefit of cross-problem pretraining. PFNs4BO is also pretrained on synthetic functions, it predicts the posterior predictive distribution of a single objective from context observations. We use the official pretrained PFNs4BO model to predict each objective independently, then apply Tchebycheff scalarization to the objective-wise LCBs to construct a preference-conditioned acquisition utility. Optimizing this utility separately for every query preference is computationally expensive. We therefore densely sample candidate designs from the decision space (1024 samples) and select the candidate with the best utility for each preference. This procedure maps the offline context and query preference to a solution through candidate scoring and selection, without problem-specific model training. The results above show that all three ICPSL variants achieve better overall PSQ ranks than PFNs4BO, supporting the benefit of generating a reusable preference-conditioned solution model directly from the offline context.

[1] Müller, Samuel, et al. Pfns4bo: In-context learning for bayesian optimization.

3. Generalization analysis.

Thank you for raising these concerns. We examine generalization under prior–target mismatch and OOD dimensionalities, and analyze how performance changes with data coverage.

3.1 Prior mismatch.

To examine transfer to problems that differ from our training prior, we construct a controlled extreme example, the Weierstrass–Takagi multi-objective problem (WTMOP), using Weierstrass [2] and Takagi [3] functions as its two objectives. These functions are continuous but extremely non-smooth, deliberately violating the smoothness of the mixed kernels during training. We use the domain [0,0.5] to ensure clear conflict between the two objectives.

Moreover, the functions have a strict Hölder exponent guarantee with controllable parameters. The Hölder exponent α\alpha measures the local roughness of a function [4,5]. A larger α\alpha indicates a smoother function. For Weierstrass, the α=−log⁡alog⁡b\alpha = -\frac{\log a}{\log b}, we set a=0.7a=0.7, b=3b=3 and α=0.325\alpha=0.325. For Takagi, α=1.0\alpha=1.0. The sample paths of a Matern-ν\nu GP have almost-Hölder exponent ν\nu. The table below compares these Hölder exponents, highlighting the smoothness mismatch between the Weierstrass-Takagi functions and our pretraining distribution.

Function / KernelWeierstrassTakagiMatérn-3/2Matérn-5/2RBF
α\alpha0.3251.01.52.5∞\infty

We evaluate the pretrained model in our paper using ten-observation contexts. The table reports PSQ ranks under LHS, Sobol, and uniform sampling.

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow5.005.006.005.330/30/3
PSLR8.008.008.008.000/30/3
PSLO7.007.007.007.000/30/3
DOMOO6.006.004.005.330/30/3
PFNs4BO1.002.001.001.332/33/3
ICPSL-MAP4.004.005.004.330/30/3
ICPSL-CAM2.001.002.001.671/33/3
ICPSL-MM3.003.003.003.000/33/3

The results show that ICPSL retains competitive preference-conditioned solution quality under this pronounced prior mismatch. CAM outperforms ParetoFlow, PSLR, PSLO, and DOMOO under every context distribution, achieving the best PSQ under Sobol sampling and the second-best under LHS and uniform sampling. MM consistently ranks third. This extreme example suggests that the model can transfer useful information learned during GP pretraining to objectives with substantially different regularity.

3.2 OOD dimension.

We evaluate generalization beyond the dimensionalities sampled during pretraining under “Generalization to OOD Dimensions” in Section 4.3 of our paper. This analysis covers both decision and objective dimensionalities. We find that ICPSL-MM achieves the best results on all tested problems with unseen decision dimensionalities, whereas ParetoFlow performs best on the problem with an unseen objective dimensionality. These results suggest that generalizing to an unseen number of objectives is more challenging in these tests than extending to an additional decision variable.

3.3 Data coverage.

We examine the effect of offline data availability under “Sensitivity to Context Size” in Section 4.3 of our paper. The results indicate that solution quality generally improves as the context grows, but the gains vary across benchmarks and inference strategies. It suggests that additional observations can provide useful problem-specific information, although their benefits are not uniformly concentrated at small context sizes. The smallest evaluation contexts contain fewer observations than those used during pretraining, probing adaptation to greater data scarcity. Continued gains with larger contexts indicate that additional observations can still refine the generated model.

We hope that the controlled prior-mismatch experiment, the OOD-dimensional evaluation, and the data coverage analysis help clarify the generalization ability of ICPSL across changes in objective-function structure, problem dimensionality, and offline data availability.

[2] Hardy, Godefroy Harold. Weierstrass’s non-differentiable function.

[3] Takagi, Teiji. A simple example of the continuous function without derivative.

[4] Evans, Lawrence C. Partial differential equations.

[5] Gilbarg, David, et al. Elliptic partial differential equations of second order.

Reviewer_8s3LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies preference-conditioned Pareto set learning from limited offline data. It proposes ICPSL, which uses a pretrained hypernetwork to construct a reusable, preference-conditioned solution distribution from a small context without problem-specific retraining. The paper derives a free-energy training objective and its KL interpretation. It conducts experiments on synthetic and engineering-design benchmarks using ten-observation contexts, reporting competitive hypervolume and faster solution-model construction.

Critical strengths
  • Amortizing both problem adaptation and preference queries could reduce the cost of repeated optimization.
  • The mixture representation can capture separated solution regions, providing flexibility at the distribution-modeling stage.
Critical weaknesses
  1. The offline information budget is ambiguous. Appendix D uses both a normalized input context and ranges estimated from 10,000 objective evaluations. It is unclear whether those ranges reach the model or only the evaluator; only the former would compromise the ten-observation setting. (pp. 17, 19-20)

  2. CAM/MM can destroy multimodal solution quality. For f1(x)=f2(x)=((x−1/4)(x−3/4))2f_1(x)=f_2(x)=((x-1/4)(x-3/4))^2, averaging the two zero-loss optima yields loss 1/2561/256. Accurate density fitting therefore does not ensure a good mean decision. (pp. 6-7, 14)

  3. Hypervolume does not validate preference alignment. Reassigning the same outputs among query preferences leaves hypervolume unchanged, so the reported metric cannot establish query-specific solution quality. (pp. 8, 20)

Questions for authors
  1. What characteristics of a target problem indicate that the synthetic pretraining prior is a poor match?

  2. How should the entropy temperature be selected as the amount of context information changes?

4 comments

Rebuttal provided very late. I will need more time to evaluate to update my score.

We thank the reviewer for the constructive comments and suggestions. We respond to each point below.

1. Normalization and offline information budget.

Thank you for requesting this clarification. The ranges estimated from 10,000 objective evaluations are used to normalize the objective values in the input context, so they do influence the model inputs. We use a fixed Sobol design to estimate objective scales over the decision domain, accounting for substantial scale differences, particularly on RE problems, and reducing dependence on the particular sparse context sampled. This follows the task-level normalization convention used in Offline Multi-Objective Optimization [1] and DOMOO [2], where normalization statistics are computed at the dataset level rather than from the current input context.

The 10,000 evaluated pairs are not supplied as additional context observations, but their extrema provide additional scale information. Our reported setting therefore uses ten context observations together with fixed benchmark-level normalization ranges, shared across all methods. We will revise the description to make both the context observation budget and this additional normalization information explicit.

[1] Xue, Ke, et al. Offline multi-objective optimization.

[2] Zhu, Yiyi, et al. Diversity-Driven Offline Multi-Objective Optimization via Nested Pareto Set Learning.

2. Solution quality under CAM/MM averaging.

Thank you for providing this example. We agree that accurately representing separated optima does not guarantee that their average preserves solution quality. Our mixture formulation is motivated by changes in the optimal solution as the preference varies. A deterministic continuous solution network may interpolate between separated solution branches, whereas a mixture can represent the transition by redistributing probability mass between low-energy regions.

During training, our objective evaluates the Tchebycheff energy of individual samples rather than their average, allowing the model to retain competing low-energy regions without concentrating probability mass at intermediate decisions.

CAM and MM provide practical strategies for extracting a single solution without further optimization. Averaging provides a representative decision when the distribution concentrates within one solution region, while quality preservation across separated regions is not guaranteed in general, as your example illustrates. Our sample-based MAP strategy offers an alternative that selects an individual candidate without cross-component averaging. We will clarify this distinction between the flexibility of the mixture representation and the properties of its single-solution inference strategies.

3. Preference-aligned performance metric.

Thank you for the helpful suggestion. We agree that HV alone does not establish query-specific solution quality. We therefore propose preference-conditioned solution quality (PSQ), defined as the average Tchebycheff value of each predicted solution under its corresponding query preference: PSQ=1T∑q=1Tmax⁡j=1,…,mλqjfj(xq)\mathrm{PSQ}=\frac{1}{T}\sum_{q=1}^{T}\max_{j=1,\ldots,m}\lambda_{qj}f_j(x_q), where TT is the number of query preferences and xq=x^(λq)x_q=\hat{x}(\lambda_q). Lower PSQ indicates better preference-conditioned solution quality. Compared with HV, PSQ explicitly evaluates each solution together with its corresponding preference.

The table reports average PSQ ranks over 30 problem–sampler cases, with lower ranks indicating better performance.

AlgorithmLHSSobolUniformOverallWinsTop-3 cases
ParetoFlow3.502.553.903.324/3016/30
PSLR6.106.006.606.231/302/30
PSLO5.305.605.505.471/306/30
DOMOO5.405.005.505.303/309/30
PFNs4BO4.105.204.304.537/3012/30
ICPSL-MAP4.254.503.454.071/3011/30
ICPSL-CAM3.853.803.553.730/3017/30
ICPSL-MM3.503.353.203.3514/3018/30

ICPSL-MM achieves an overall average rank close to ParetoFlow while obtaining the most wins and Top-3 cases. These results provide additional evidence of competitive preference-conditioned solution quality beyond the set-level assessment offered by HV. We will include this evaluation to support our claims about preference-aligned performance.

4. Characteristics of prior–target mismatch.

Thank you for this question. In our setting, where objective values are normalized, potential indicators of a poor prior match include substantially lower smoothness or shorter length scales outside the range represented during pretraining. Lower smoothness may appear as sharp kinks or irregular local behavior, while shorter length scales may produce rapid oscillations or narrow low-energy regions between nearby observations. Such features can be underrepresented by our prior and difficult to infer from a sparse context.

We examine both properties experimentally. The smoothness stress test evaluates transfer to highly irregular objective landscapes, where ICPSL-CAM achieves the best results under Sobol sampling and the second-best under LHS and uniform sampling. This shows that substantial differences in objective regularity can still permit competitive transfer. Detailed experimental results and analysis are provided in Section 3.1 of our response to Reviewer AgHE.

Our length-scale ablation further shows that pretraining with shorter length scales worsens solution quality under all three inference strategies, whereas longer length scales improve it on the evaluated benchmarks. These results suggest that emphasizing longer-range correlations during pretraining provides a more useful inductive bias for the evaluated benchmarks. Detailed results and analysis are provided in Section 2 of our response to Reviewer dKui.

5. Temperature selection across context sizes.

Thank you for this question. Context information determines how much evidence the model has to infer solution energies, while the temperature controls how strongly energy differences translate into differences in probability density. A more informative context could support a smaller temperature, placing greater emphasis on lower-energy solutions. However, context size alone does not establish an optimal temperature schedule. We use τ=0.015\tau=0.015 during pretraining and retain the same pretrained model across context sizes, allowing the generated distribution to adapt through the context input. The context-size analysis in Section 4.3 of our paper evaluates this fixed-temperature setting.

Dear Reviewer 8s3L,

Thank you for your constructive comments and suggestions. We apologize for the delay in submitting our rebuttal. We have now submitted our responses and hope they address your concerns.

Thank you again for your time and understanding.

Best regards,

The Authors of Paper 43330

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your constructive comments and suggestions. We are currently waiting for the results of additional experiments, as some ablation studies require multiple rounds of model pretraining and take considerable time to complete. We sincerely appreciate your understanding and will submit our rebuttal responses as soon as possible.

Thank you again for your help and understanding during the discussion process.

Best regards,
The Authors of Paper 43330

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.