acceptodds
Under review as a conference paper at ICLR 2027

From Fixed Archives to New Operating Conditions: Offline Parametric Multi-Objective Optimization

Abstract

Many engineering systems must operate under varying conditions, such as turbines facing different wind speeds or vehicles carrying different payloads, and the designs that are best for one condition are rarely best for another. Because cost, safety, and performance must be balanced at once, each condition poses its own multi-objective problem with its own Pareto set, and repeating the optimization for every new condition is too slow for timely deployment. In practice, substantial historical evaluation data are often available, which motivates reusing them offline to reconstruct trade-off solutions for previously unobserved conditions without new evaluations. The challenge in this offline parametric multi-objective optimization (Offline PMOO) setting is to generalize to unseen conditions and to recover diverse trade-offs without exploiting model errors in regions with sparse data, which neither current offline optimization nor parametric multi-objective methods address. We therefore formalize Offline PMOO and construct a benchmark of 27 tasks, ranging from synthetic test functions to engineering applications. After investigating the limitations of task-conditioned surrogate models in this regime, we propose a generative refinement strategy, instantiated via Parametric Scalarized Diffusion (PSD) and Parametric Pareto Flow Matching (PPFM), which learns a conditional prior from archived elite designs and uses surrogate models purely to steer refinement, so that candidates do not drift into unsupported regions. PPFM and PSD rank first and second on the synthetic and shared-component engineering suites. Our evaluation also maps the limits of learned solvers: surrogate-only models are preferable on a conditional simulator suite, and zero-learning archive reuse suffices for many-objective problems where Pareto dominance loses its discriminative power.

Primary area:
optimization
Keywords:
offline multi-objective optimization, parametric multi-objective optimization, Pareto set learning, diffusion models, flow matching

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

This paper introduces Offline Parametric Multi-Objective Optimization (Offline PMOO), where a method must reconstruct a Pareto set for a previously unseen operating condition using only an archive collected at other conditions, without additional oracle evaluations. The paper formalizes the setting, states applicability assumptions, provides an unidentifiability result for unrestricted zero-query optimization, and proposes a benchmark containing 27 synthetic and engineering instances.

The proposed methods, Parametric Scalarized Diffusion (PSD) and Parametric Pareto Flow Matching (PPFM), learn a task-conditioned generative prior from the nondominated portion of each archived task slice. At inference time, candidates are initialized from a nearby archive slice, improved through gradient descent on a learned surrogate scalarization, and refined with guided diffusion or flow matching. The experiments compare these methods with several task-conditioned surrogate models, nearest-archive reuse, a warm-start-only method, and a modified parametric Pareto-set-learning baseline.

Critical strengths
  1. Important and clearly motivated problem setting. The paper identifies a practically relevant gap between offline optimization and parametric multi-objective optimization. The zero-query deployment regime is clearly distinguished from contextual Bayesian optimization, dynamic MOO, and standard offline MOO.
  2. Thoughtful problem formulation. The applicability conditions A1–A4 are useful for understanding when Offline PMOO can be expected to work. Even though the theoretical results are limited, explicitly connecting archive coverage, task distance, surrogate reliability, and support displacement is conceptually helpful.
  3. Sensible algorithmic design. Combining an archive-supported initialization with a learned elite prior is a natural way to reduce surrogate exploitation. The division of labor between the surrogate, which provides preference-specific directions, and the generator, which regularizes candidates toward supported designs, is well motivated.
  4. Broad and unusually candid empirical analysis. The paper evaluates 27 instances across three suites and reports several informative diagnostics. In particular, it openly identifies regimes where a surrogate solver or even nearest-archive reuse is preferable. The archive-composition, parameter-sparsity, surrogate-exploitation, and inference-time sensitivity studies are valuable.
  5. Well presentation and implementation detail. The paper is generally well organized, and the appendix contains enough architectural, optimization, and data-collection information to make reproduction plausible.
Critical weaknesses
  1. The main comparison uses unequal candidate-set sizes. The main Table 2 comparison is not cardinality-matched: PSD and PPFM return K=256K=256 candidates, whereas the surrogate solvers return an NSGA-II population of 100. Since hypervolume is monotone with respect to adding nondominated candidates, this confounds method quality with output-set cardinality. Appendix I.2 is useful, but it covers only four selected tasks and shows that rankings can change with KK, most notably on DF10. The main 27-instance ranking should therefore be recomputed at a common candidate budget.
  2. The baseline coverage remains incomplete. The paper discusses ParetoFlow, Preference-Guided Diffusion, and Pareto-Conditioned Diffusion, but does not evaluate natural task-conditioned adaptations of these closely related offline-MOO generative methods. In addition, the PPSL-MOBO baseline replaces the original hypernetwork LoRA preference-to-design map with concatenation. This may be a reasonable simplified variant, but results from HN-LoRA surrogate predictors do not establish that the corresponding PPSL parameterization would behave similarly. These omissions weaken claims about the benefit of the specific PSD/PPFM construction relative to prior generative approaches.
  3. The uncertainty reporting for the main ranking is incomplete. Table 2 reports only average ranks, while Appendix J averages EHV over seeds and reports variation primarily across task parameters. This does not quantify training-seed uncertainty, and task parameters evaluated by the same trained model should not be treated as independent replicates. There is also an apparent inconsistency: Appendix C.2 describes ranking separately for each seed and task parameter, whereas Appendix J describes ranking seed-averaged hypervolumes and states that this reproduces Table 2. These two procedures are generally not equivalent. The authors should clarify the aggregation rule and report seed- or problem-level confidence intervals, preferably using paired or hierarchical resampling.
Questions for authors
  1. How do the Table 2 rankings change when every method returns exactly the same number of candidates, e.g., K=100K=100, on all 27 benchmark instances?
  2. Do the advantages of PSD and PPFM persist under a matched inference budget, such as the same number of surrogate/gradient evaluations or comparable wall-clock time, in addition to matched output cardinality?
  3. Can the authors compare against task-conditioned adaptations of closely related offline-MOO generative methods? This would help isolate the benefit of the proposed archive-supported refinement design.
  4. What is the exact aggregation procedure used for Table 2? Appendix C.2 appears to rank methods separately for each seed and task parameter, whereas Appendix J describes ranking seed-averaged hyper volumes. Since these procedures are not generally equivalent, can the authors clarify the implementation and report seed- or problem-level confidence intervals for the main comparisons?
Reviewer_t6aDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper tries to address how to reconstruct the Pareto set of a multi-objective problem at a new, unseen task parameter (an operating condition such as wind speed or payload) using only a fixed archive of designs evaluated at other task parameters, without any new evaluation. The authors call this setting offline parametric multi-objective optimization (Offline PMOO), state conditions under which it is well posed, and prove that no zero-query method has an unconditional guarantee. They build a benchmark of 27 instances (14 synthetic, 9 engineering, 4 conditional simulators), study 11 task-conditioned surrogate solvers, and propose a warm-started generative refinement, instantiated with diffusion (PSD) and flow matching (PPFM), which starts from archived designs, takes surrogate gradient steps, and then partially noises and denoises the design with a generator trained on archived elite designs. They find that PPFM and PSD have the best average ranks on the synthetic and engineering suites, while a surrogate solver ranks first on the simulator suite and nearest-archive reuse ranks first on nine-objective problems.

Critical strengths

I think the setting is practical and clearly stated. The motivating examples (behaviour repertoires for robots, soft robots in different environments, components used under varying conditions, p. 1) are real, and the conditions under which the problem is well posed are spelled out (A1-A4, p. 3). The empirical study is broad and the analysis is careful. The benchmark covers 27 instances with a held-out-parameter protocol and the code is released (p. 55). The authors also analyze surrogate exploitation and archive coverage (Tables 15, 18 and 21) and openly report where their methods do not win, even advising practitioners to "run the zero-learning control before adopting a more complex method" (p. 20).

Critical weaknesses
  1. I am not convinced that Offline PMOO needs to be a new problem rather than conditional offline MOO. Offline MOO is established, and so is parametric MOO. If the task parameter is treated as a conditioning input that is fixed at test time, the problem becomes offline MOO in which only the design variables are optimized. The paper's own surrogate solvers do exactly this (for example, Cat concatenates the task parameter with the design and runs NSGA-II at the queried parameter, p. 5), and one of them, Cat-E2E-G, ranks first on the conditional simulator suite (6.27 vs. 6.97 for PSD and 7.91 for PPFM, Table 2, p. 7). The only argument against reducing the problem to offline MOO (Appendix A.4, p. 20) is that per-parameter offline MOO cannot handle unseen parameters, which does not address conditioning on the parameter. The paper does not mention interpretability or any other reason why the task parameter must be treated separately. I think the authors need to explain rigorously what Offline PMOO requires that a conditional version of offline MOO cannot provide.

  2. The most direct baselines from offline MOO are missing. For a paper that combines offline MOO and parametric MOO, the main comparisons cover only one side. Generative offline MOO methods, namely ParetoFlow (Yuan et al., 2025a), Preference-Guided Diffusion (Annadani et al., 2025) and Pareto-Conditioned Diffusion (Shrestha et al., 2026), are described in the related work (p. 18) but not run, and the paper gives no reason for this. The same holds for design editing (Yuan et al., 2025b), which the paper describes as applying noise injection and denoising to gradient-optimized designs (p. 17) and which is the closest prior method to PSD, and for the generative parametric MOO method of Wei et al. (2025). The standard offline MOO benchmark, Off-MOO-Bench (Xue et al., ICML 2024), is not cited, although it already includes RE problems and adapts existing methods along the data, model architecture, learning algorithm and search algorithm, which overlaps with the surrogate study here. Conditional versions of these generative methods are the natural competitors for PSD and PPFM, and I think these experiments are necessary.

Other weaknesses:

  • The statistical evidence is thin. The number of seeds for the main synthetic and engineering experiments is not stated, and no confidence intervals or significance tests are reported; the main table gives only average ranks. Per problem, PSD or PPFM ranks first on only 6 of the 14 synthetic problems (Table 32). The generative solvers return 256 designs while the surrogate solvers return 100 (p. 35), which favours the former under hypervolume.
  • The method has limited novelty, and the ablation does not support a single coherent method. The refinement is essentially a task-conditioned version of noise-and-denoise after gradient optimization. In Table 17, guidance and elite selection change PSD by less than 1% on three of four tasks, and removing PSD's warm start improves DF4 by 14.4% and RE34-S1 by 6.3%. The ablation also uses an earlier, smaller held-out set and a two-front elite setting (pp. 24, 45).
  • The theory is generic and imprecisely stated. Theorem 1 is the standard construction that adds a bump function at an unseen parameter; it says that nothing can be guaranteed where no data are observed, which holds for any extrapolation problem and is not specific to Pareto sets. The main text says the Pareto sets can be "arbitrarily far apart" (p. 3), while the proof only gives a separation of the order of the diameter of the bounded design space, and it first takes F(1) to be "any" consistent family but later chooses it so that its Pareto set lies near a specific design (p. 22). The constants in Proposition 1 are not estimated (p. 23).
  • The text and tables disagree in several places. On DF12, the text says the generative solvers stay near 1.00 while the surrogate oscillates between 0.89 and 0.95 (p. 8), but in Tables 31 and 32 the surrogate solver Cat-E2E-P is best on DF12 (0.842 vs. 0.693 and 0.646). Appendix G.2 claims the advantage over surrogate-based methods is largest at m = 4 (p. 39), but in Table 2 Warm-Start-Only ranks better (2.56) than PPFM (2.62) and PSD (3.04). The cost of an elite-only archive to a surrogate baseline is given as up to 25% (pp. 20, 26) and as 85% (p. 9, Table 18).
Questions for authors
  1. Why can Offline PMOO not be solved as conditional offline MOO, with the task parameter as a conditioning input that is fixed at test time? What does the new setting require that this reduction cannot provide?
  2. How do PSD and PPFM compare with task-conditioned versions of ParetoFlow, Preference-Guided Diffusion and Pareto-Conditioned Diffusion, and with design editing (Yuan et al., 2025b)?
  3. How many seeds are used for the main synthetic and engineering experiments, and are the differences between the top methods statistically significant?
Reviewer_bLehOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates Offline Parametric Multi-Objective Optimization (Offline PMOO), where Pareto-optimal solutions must be generated for unseen operating conditions using a fixed historical archive without additional objective evaluations. It constructs a benchmark comprising 27 tasks and proposes two generative refinement methods, Parametric Scalarized Diffusion (PSD) and Parametric Pareto Flow Matching (PPFM), which combine conditional generation with surrogate-guided optimization. Experiments demonstrate promising performance on synthetic and shared-component engineering tasks, while also revealing limitations on real-world conditional design tasks and many-objective problems.

Critical strengths
  1. Offline PMOO addresses a practically meaningful challenge in multi-objective optimization under limited evaluation budgets.
  2. The evaluation covers 27 tasks across synthetic, engineering, and real-world conditional design settings.
  3. Combining generative priors with surrogate-guided refinement offers a plausible way to mitigate surrogate exploitation.
Critical weaknesses
  1. The proposed methods appear to combine existing generative modeling and surrogate-guided optimization techniques. Their methodological advances over existing generative multi-objective optimization approaches require clearer justification.
  2. Direct comparisons with representative existing generative offline multi-objective optimization methods are limited, making it difficult to establish the proposed methods' distinct advantages.
  3. Ablation studies on a limited number of representative tasks do not fully establish the general effectiveness of generative refinement across different problem characteristics.
  4. The normalization procedure for REAL tasks depends on candidate solutions produced by the evaluated methods, potentially affecting comparability. Further sensitivity analyses and statistical significance tests would strengthen the conclusions.
Questions for authors
  1. What are the fundamental algorithmic innovations of PSD and PPFM beyond combining existing generative models with surrogate-guided refinement?
  2. How sensitive are the REAL-suite results and method rankings to the objective normalization bounds and hypervolume reference points?
  3. Under what conditions does generative refinement provide consistent benefits over surrogate-only optimization or archive reuse, and can these conditions be identified from archive coverage or task characteristics?
  4. Statistical significance tests are needed to establish whether the observed performance differences across methods are robust.

Comments on the paper

1 comment

Dear Area Chairs,

Thank you very much for your time and effort in handling our submission (#36525). As the discussion period concludes, we would like to briefly note the discussion status and summarize our revisions and core contributions to assist your meta-review.

Following our rebuttal, only Reviewer t6aD participated in the discussion, confirming that our new experiments resolved their evaluation concerns and raising Soundness from 2 to 3. Reviewers bLeh and ZFfg did not respond during the discussion phase. Because their primary critiques focused on concrete experimental and statistical requests that are now fully addressed in the revised manuscript, we respectfully hope you will take these updates into account when assessing our paper's contributions.

Summary of Key Revisions

  1. Generative baselines and inverse-model comparisons (Reviewers bLeh, t6aD, ZFfg): We added five task-conditioned generative MOO baselines—ParetoFlow, DEMO (design editing), PGD-MOO, PCD, and Wei et al. (2025)—under an equal budget of K=100K = 100 across all 24 deterministic settings (Table 3, Appendix F.5, Table 52). PSD and PPFM outperform all five generative baselines on 22/24 to 24/24 settings (pHolm≤1.79×10−4p_{\text{Holm}} \le 1.79 \times 10^{-4}) and beat the per-setting best-external-mean envelope on 19/24 and 21/24 settings. We also added a matched 9-setting comparison between Concat PSL and the HN-LoRA inverse map (Appendix H.2, Table 21).

  2. Unified candidate budget and inference-cost evaluation (Reviewers t6aD, ZFfg): We recomputed all 16 methods at a matched budget of K≤100K \le 100 across 5 archive replicates and 16 unseen conditions per setting (Table 2, Appendix J.1), where PPFM ranks 1st overall and PSD ranks 4th (ranking 1st and 2nd on Synthetic in HV, IGD+, and IGD, and top two on REAL). We also added wall-clock latency and evaluation-count caps over 960 serial runs (Table 4b, Figure 2, Appendix I): under a 0.50 s cap, PPFM leads on DF4, DF10, and Solar, whereas running NSGA-II longer degrades true EHV on DF10 (0.6189→0.40230.6189 \to 0.4023) and Solar (0.7697→0.70940.7697 \to 0.7094) due to surrogate exploitation.

  3. Statistical rigor, aggregation clarity, and metric sensitivity (Reviewers bLeh, t6aD, ZFfg): Using the 5 archives as paired resampling units, we added archive-level SDs, bootstrap rank CIs, and paired CIs with exact tests across all 432 contrasts, alongside Holm-corrected benchmark sign tests (Tables 2, 44–51, 57–66). We reported both rank aggregation rules (PPFM ranks 1st under both; Table 51) and audited 8 normalization and reference-point schemes on REAL, where 38/40 external comparisons and 100% of control comparisons retain their signs (Appendix G.7).

  4. Problem positioning, theory, and expanded ablations (Reviewers bLeh, t6aD): We explicitly formulated Offline PMOO as conditional offline MOO evaluated on unseen operating conditions (citing Off-MOO-Bench; Sections 1–3, Appendix A), tightened Theorem 1 and Proposition 1 (Appendix B), and expanded component ablations to 8 settings alongside 24-setting paired controls (Generator-removed, Selected nongenerative, and Same-guide NSGA-II; Table 4a, Appendix H.4, Tables 53–56).

Core Contributions and Methodological Takeaway

Regarding Reviewer t6aD’s remaining hesitation on novelty and RE-suite gains: our work provides a zero-query unseen-condition evaluation protocol, a 24-setting benchmark, and a systematic study of conditional surrogates, inverse maps, and warm-started generative refinement. Beyond substantially outperforming standalone generative editing (DEMO) and four other generative MOO methods, PPFM improves over its exact generator-removed control on 21/24 settings. Crucially, our study characterizes when direct surrogate search suffices versus when generative refinement is essential: on low-dimensional, smooth engineering surfaces (the 2–6 variable RE tasks), surrogate error stays small away from the archive (RMSE 0.0065–0.0133) so direct NSGA-II excels; on higher-dimensional or sharply deforming landscapes (Synthetic and REAL), unconstrained search exploits model errors while archive-anchored refinement prevents degradation.

Thank you again for your time and consideration.

Best regards,

The Authors of Submission 36525

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.