acceptodds
Under review as a conference paper at ICLR 2027

ParetoTransport: Generative Optimization by Mass Transport Toward The Pareto Front

Abstract

Offline multi-objective optimization (MOO) seeks candidate designs whose objective vectors not only lie close to the Pareto front but are also effectively distributed along it. Generative methods have recently emerged as a promising approach because they learn a distribution over feasible designs while allowing generation to be guided toward desirable designs. Existing methods, however, largely retain classical sample-wise guidance strategies, leaving the distribution-level modeling capability of generative methods underused. We formulate offline MOO as a sampling problem and motivate the -Wasserstein distance as a natural quality measure. Building on this view, we propose ParetoTransport, a training-free guidance method for pre-trained flow-matching models that formulates guidance directly over the population-level distribution in objective space, combining directional transport toward the Pareto front with Wasserstein matching that controls the allocation of probability mass along it. Under an exact-matching assumption, we prove that the transported distribution converges geometrically toward the Pareto front in -Wasserstein distance. On offline MOO benchmarks, we extend the evaluation of recent generative methods beyond hypervolume to generational distance, inverted generational distance, and the -Wasserstein distance. ParetoTransport achieves consistently strong performance, attaining the best average rank in convergence and probability-mass allocation across all benchmark families.

Primary area:
generative models
Keywords:
Offline multi-objective optimization, flow matching, training-free guidance

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eKTUOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper describes ParetoTransport, a training free distribution level guidance technique that explicitly considers the objective space performance. The authors model the offline MOO problem as a statistical sampling problem, transforming the goal of MOO to finding a distribution in the design-space, such that the induced objective space distribution is uniformly allocated across the unknown Pareto-Front. To measure the quality, the authors propose to use 2-Wasserstein distance, which is motivated by its relation to GD and IGD metrics. Compared to existing methods, the authors directly formulate the guidance over the probability path. During the sampling, the authors perform the optimization directly in terminal space, avoiding constructing optimistic estimates in the objective space. This results in an objective that considers both the W2 probability-mass allocation and the direction of the sampling. Since the true Pareto-Front is unknown, the authors use CHIM normal as an approximation of the projection direction.

To validate the proposed approach, the authors prove that under certain conditions, their approach converges towards Pareto-Front in 2-Wasserstein distance. Empirically, they consider synthetic and RE benchmark suites from offline-MOO benchmark, measuring the performance with HV, GD, IGD and W2 distance. The results showcase consistent performance in all metrics, either outperforming or matching other methods. Furthermore, the authors validate that both the W2 term and directional term are required for consistent performance.

Critical strengths
  • The authors propose a well motivated, training free guidance technique that considers the distribution level performance in the objective space. To achieve this, they introduce a sequence of intermediate proxy distributions, alternating between performing W2 matching and performing directional transport. I believe this approach brings a new view to the training-free offline MOO methods, as it directly considers the distribution level guidance, instead of relying on sample-wise guidance terms, as done previously.
  • The authors provide a theoretical proof, that under reasonable assumptions, shows that the proposed method geometrically converges towards the Pareto Front.
Critical weaknesses
  • The authors motivate the use of 2-Wasserstein distance as a quality metric from the viewpoint of GD and IGD. However, it is well known that GD and IGD are not Pareto-compliant [1], i.e. they can rank solution sets in incorrect order, making the claim that this is a natural quality metric questionable.
  • The authors' evaluations are limited to synthetic and real-world design tasks, while e.g. ParetoFlow [2] and PCD [3], utilize a much larger set of benchmarks in their evaluations.
Questions for authors
  • ParetoFlow [2] and PCD [3] also consider MORL, MONAS and SciDesign tasks, which provide a more comprehensive performance profile. Given that you share the network structure with ParetoFlow, I find the omission of these tasks surprising. Are there any specific reasons for these omissions?
  • You start your mass-transport from the empirical distribution D(best)D(\mathrm{best}). While this is quite understandable, I would be interested in seeing what happens if the initial distribution u0u_0 is not ideal, e.g. it has very narrow coverage or it is far away from the Pareto Front?
  • Currently, the proxy update operator assigns probability masses either using uniform distribution or Crowding distance. Did you consider using e.g. hypervolume contribution [4] or any other metrics that are directly related to Pareto dominance?
  • It would be nice to see some examples of the generated solution sets in e.g. synthetic functions, where one can see how the method performs when the solution set is e.g. non-convex.
  • You perform evaluation using GD and IGD. However, it is well known that both of these metrics may rank solution sets incorrectly (and thus there exists IGD+ and GD+ variants). If you have the resulting solution sets saved, I would suggest using the + variants of these metrics for more robust evaluation. (NOTE: I am not asking you to perform all experiments again).

[1] Ishibuchi, H., Masuda, H., Tanigaki, Y., & Nojima, Y. (2015). Modified Distance Calculation in Generational Distance and Inverted Generational Distance. In Evolutionary Multi-Criterion Optimization (pp. 110–125).

[2] Yuan, Y., Chen, C., Pal, C., & Liu, X. (2025). ParetoFlow: Guided Flows in Multi-Objective Optimization. The Thirteenth International Conference on Learning Representations.

[3] Shrestha, J., Heiskanen, S., Hepola, K., Rissanen, S., Jääskeläinen, P., & Pajarinen, J. (2026) Pareto-Conditioned Diffusion Models for Offline Multi-Objective Optimization. The Fourteenth International Conference on Learning Representations.

[4] Bringmann, K., Friedrich, T. (2009). Approximating the Least Hypervolume Contributor: NP-Hard in General, But Fast in Practice. In Evolutionary Multi-Criterion Optimization. EMO 2009. Lecture Notes in Computer Science, vol 5467

2 comments

We thank the reviewer for their positive assessment. We are glad the reviewer finds the method well motivated, bringing a new view to training-free offline MOO, with geometric convergence in theory and consistent performance in all metrics. We address each weakness and question below.

Pareto-compliance and Q5 We agree that GD and IGD are not Pareto-compliant. However, the bound of Corollary 1 also extends to the modified variants GD+GD^{+} and IGD+IGD^{+} [1]: d+(a,z)≤∥a−z∥d^{+}(a, z) \leq \lVert a - z\rVert and thus GD+≤GD≤W2GD^{+} \leq GD \leq W_2 and IGD+≤IGD≤W2IGD^{+} \leq IGD \leq W_2. We have added new evaluations (Tables 47-53 in Appendix A.7) of GD+GD^{+} and IGD+IGD^{+}. ParetoTransport remains first in five of the six columns under GD+ and IGD+, and moves from second to third in IGD+ on RE. We agree that W2W_2 is not Pareto-compliant either and should be seen as a novel complement: it is natural for generative models operating on probability distributions and additionally measures the allocation of probability mass along the front.

Table 47. Avg. rank, GD→GD+ and IGD→IGD+. Complete table in Appendix A.7.

MethodZDT GDZDT IGDDTLZ GDDTLZ IGDRE GDRE IGD
MultiHead-Vanilla4.000→4.0005.400→4.0005.929→5.9295.214→5.3576.375→6.5835.833→5.125
MultipleModels-Vanilla3.800→3.8005.600→5.0004.143→4.1435.714→5.3575.708→5.7085.333→3.833
ParetoFlow8.000→8.0009.600→9.2008.429→8.4298.571→8.8576.042→6.2086.917→8.875
PGD9.200→9.2006.200→6.6007.214→7.2147.000→6.7148.417→7.8336.000→7.250
ParetoTransport (ours)3.600→3.6002.000→1.8001.571→1.5712.000→1.7863.875→3.5835.083→4.542

Benchmark and Q1. We use the standard benchmarks of offline MOO, ZDT [5], DTLZ [6] and RE [7], which recent generative approaches [8, 9] also use. We initially focused on these tasks because the benchmark provides reference Pareto fronts for them, which GD, IGD and W2W_2 require. Following the reviewer’s suggestion, we have added the 13 MO-NAS tasks with m≤3m \le 3 objectives (Tables 43–46 in Appendix A.6), using the non-dominated designs of the held-out test split as reference set: ParetoTransport attains the best average rank in HV, IGD and W2W_2 and the second best in GD. We leave MORL and Sci-Design to future work.

Table 2. Avg. rank on MO-NAS. Complete table in Section 7.1.

MethodHVGDIGDW2W_2
MultipleModels-COM6.0776.9235.5384.462‾\underline{4.462}
MultipleModels-ICT6.6153.8466.0005.846
ParetoFlow-Vanilla5.5386.0004.7695.692
PGD2.385‾\underline{2.385}6.4622.308‾\underline{2.308}5.538
ParetoTransport (ours)1.3854.846‾\underline{4.846}1.3853.231

Q2 Non-ideal initial distribution.

Theorem 1 distinguishes two ways in which the initial distribution can be far from the Pareto front. Far along the projection direction, W2(μ0,μˉ)W_2(\mu_0, \bar\mu), which equals the GD of the offline front (Remark 4): this is the default setting of the ZDT [5] and DTLZ [6] tasks, where the offline front has a GD of 0.262 and 0.373. ParetoTransport reduces this by 46% and 48%, respectively (Table 4 in Section 7.2).

Far in terms of the residual, W2(μˉ,μ∗)W_2(\bar\mu, \mu^\ast), as for narrow coverage: we ran a new experiment (Figure 6 in Appendix A.1.2) on RE21 with the offline front restricted to f1<0.5f_1 < 0.5. The solution set reaches the Pareto front within the covered region but does not extend beyond it. We state this as a limitation in the conclusion and leave extrapolation beyond the covered region to future work.

Q3 Hypervolume contribution [4] We appreciate this interesting suggestion. Our goal is an even allocation of mass along the front, and weighting proxy points by crowding distance empirically promotes this by shifting mass toward sparsely populated regions. Hypervolume contribution [4] is a valid alternative but does not serve this goal: even equally spaced points receive masses that vanish as the front becomes steep or flat.

Q4 Non-convex example. We have added a new Figure 5 in Appendix A.1.2, which shows the solution set with the concave Pareto front of ZDT2.

[5] Zitzler et al., Comparison of Multiobjective Evolutionary Algorithms: Empirical Results, Evolutionary Computation, 2000. [6] Deb et al., Scalable multi-objective optimization test problems, CEC, 2002. [7] Tanabe and Ishibuchi, An easy-to-use real-world multi-objective optimization problem suite, Applied Soft Computing, 2020. [8] Annadani et al., Preference-Guided Diffusion for Multi-Objective Offline Optimization, NeurIPS 2025. [9] Hotegni and Peitz, SPREAD: Sampling-based Pareto front Refinement via Efficient Adaptive Diffusion, ICLR 2026.

I thank the authors for detailed responses and well designed additional experiments. I will keep my positive assessment of this paper.

Reviewer_21HYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The work introduces an inference-time method for distributional adaptation of pre-trained generative models, to adapt them to increase coverage of the Pareto frontier in multi-objective generative (offline) optimization. The authors propose a principled, iterative method based on mass transport, and showcase experimental results on synthetic and real-world settings.

Critical strengths
  1. (Problem) The work tackles an important, timely problem, i.e., multi-objective generative optimization.
  2. (Method) The proposed method is principled and shows promising experimental results on the (limited) settings analyzed.
Critical weaknesses
  1. (Experiments) To my understanding, all the experiments proposed are very toysh. It seems to me that the highest-dimensional "real-world" task considered is 7-dimensional, while synthetic settings reach 30-dims. These experiments do not seem to assess the method in the vast majority of applications of generative modeling. Even without going to large models (eg for proteins/images), the method should at least be assessed on relevant data in a few hundred dimensional settings. It is otherwise impossible to determine if the proposed algorithm can minimally work on modern generative models. This is especially the case of distributional adaptation processes, for which typically "bad" algorithms have a bad dependency on dimensionality.
  2. (Prior Works / Scope) This work fits right into a fast-paced literature that grew over the last few years of methods that perform distributional adaptation of models. Nonetheless, the writing seems to fully neglect it, and at times it seems to implicitly argue that it is the first method that performs distributional control (e.g., lines 475-477 in Conclusion). The general problem of adapting models according to distributional functionals, for which objectives such as Expected Hypervolume are a subcase, has been arguably formalized in [1]. While prior works have tried to tackle inference-time counterparts of specific sub-problems, such as to increase sample diversity [2]. This work seems to develop an inference-time counterpart of [1] for a specific objective, which has been already previously done for other objectives [e.g., 3]. I believe it is important to clarify the relationships with prior works on distributional adaptation (both fine-tuning on objectives that include the one studied here as a subcase, and inference-time on other objectives), since the challenges and methodologies are vastly shared across different functionals (as shown in [1]).

References

[1] Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning. De Santi et al., NeurIPS 2025

[2] Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models. Corso et al., ICLR 2024.

[3] Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuning. Wang et al., ICML 2026

Questions for authors

Regarding Weakness 1, please let me know if I misinterpreted the scale of the current experiments within the paper.

7 comments

We thank the reviewer for their careful reading of our work. We are glad the reviewer recognized that the work tackles an important and timely problem with a principled method. We address each weakness and question below.

On the dimensionality and Q1. Summary. We agree that scaling with dimension is a key concern for distributional adaptation methods. Our problem setting, however, requires only computing the distributional coupling in objective space (m≤3m \le 3), independent of the design dimension dd; thus it scales with dd like sample-wise guidance [5].

Technical Details. The distributional coupling is the transport plan PP for the cost Cij=∥f(x1i)−yj∥2C_{ij} = \lVert f(x^i_1) - y^j\rVert^2 between points in Rm\mathbb R^m; its computation does not involve dd. By Danskin's theorem, the gradient is taken with PP fixed (Appendix A.3.3), and for fixed PP the loss ∑i,jPij∥f(x1i)−yj∥2\sum_{i,j} P_{ij} \lVert f(x^i_1) - y^j\rVert^2 is a sum over samples. Hence dd enters only through the surrogate ff, as in sample-wise guidance.

We follow the standard benchmarks of offline MOO, ZDT [6], DTLZ [7], and RE [8], and recent generative approaches [9, 10]. The reviewer's reading is correct: d≤30d \le 30 (synthetic) and d≤7d \le 7 (real-engineering). Following the reviewer's suggestion, we also performed new experiments (Tables 43-46) in higher dimensions and added 13 MO-NAS tasks [11], with dd = 40–195. The results are in line with the lower-dimensional ones: ParetoTransport attains the best average rank in HV, IGD and W2W_2 and the second best in GD (Table 2), and thus remains best in W2W_2 on all four benchmark families.

Changes. We have added the 13 MO-NAS tasks (Tables 43–46 in Appendix A.6).

Table 2. Avg. rank on MO-NAS. Complete table in Section 7.1.

MethodHVGDIGDW2W_2
MultipleModels-COM6.0776.9235.5384.462‾\underline{4.462}
MultipleModels-IOM6.1547.6156.4626.000
MultipleModels-ICT6.6153.8466.0005.846
ParetoFlow-Vanilla5.5386.0004.7695.692
PGD2.385‾\underline{2.385}6.4622.308‾\underline{2.308}5.538
ParetoTransport (ours)1.3854.846‾\underline{4.846}1.3853.231

[5] Yuan et al., ICLR 2025. [6] Zitzler et al., Evolutionary Computation, 2000. [7] Deb et al., CEC 2002. [8] Tanabe and Ishibuchi, Applied Soft Computing, 2020. [9] Annadani et al., NeurIPS 2025. [10] Hotegni and Peitz, ICLR 2026. [11] Xue et al., ICML 2024.

On Prior works/scope.

Summary. We thank the reviewer for pointing us to the interesting line of research building on generative optimization via fine-tuning [1, 3, 4]. We agree that ParetoTransport can be viewed in the memoryless stochastic setting as an inference-time counterpart of [1], however, the two formulations differ in our deterministic setting.

Technical Details. In the deterministic setting of our Eq. (6), the penalty for deviation from the pre-trained model depends not only on the terminal density. Eq. (6) is equivalent to min⁡qJ(q)+λCu(q)\min_q J(q) + \lambda \mathcal C_u(q), where Cu(q)\mathcal C_u(q) is the minimal kinetic energy needed to reach qq from p0p_0 under the pre-trained model uu. This is Eq. (5) in [1] with F:=−JF := -J and Cu\mathcal C_u in place of DD. In the memoryless stochastic setting, the corresponding minimal path-KL penalty equals KL(q∥p1)\text{KL}(q \Vert p_1), exactly as in [1]; in the deterministic setting of Eq. (6), Cu\mathcal C_u depends on uu not only through p1p_1.

Changes. We have (i) made lines 475–477 (now 492-495) of the conclusion more precise to remove the ambiguity:

More broadly, generative models inherently operate on distributions over designs, and a growing literature adapts them according to functionals of the generated distribution rather than functions of individual samples [1-4].

and (ii) added a paragraph to the introduction that introduces distributional adaptation [1–4]:

This form of control is the subject of a growing literature on the distributional adaptation of pre-trained generative models, that is, their adaptation according to functionals of the generated distribution rather than functions of individual samples. For training-time adaptation, it has been formalized as generative optimization via fine-tuning [1], with instantiations for tail-aware and multi-objective optimization [3, 4]. At inference time, particle guidance couples samples to promote diversity [2], without specifying the resulting distribution.

and (iii) added a paragraph on distributional adaptation to the extended related work (Appendix A.8).

[1] De Santi et al., NeurIPS, 2025.

[2] Corso et al., ICLR 2024.

[3] Wang et al., ICML 2026.

[4] Guevara et al., Multi-Objective Generative Optimization via Expected Hypervolume Fine-Tuning, arXiv:2611.05207, 2026.

Dear Reviewer 21HY,

Thank you again for your valuable feedback. If our answers and changes have addressed your concerns, please take this into account in your final score. We are happy to engage in further discussion.

Best regards,

The Authors

Reviewer_21HYOpenReview ↗

Updated pdf?

The authors keep referring to an updated pdf, but I cannot see the updated manuscript. I see only one (the original) version. Are the authors sure that they have indeed updated the manuscript? E.g., I cannot find neither the reported new experimental results / Tables 43-46, nor the updated extended related works section (Appendix A.8 is fully missing).

Reviewer_a8fqOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

I understand ParetoTransport as a guidance method for pretrained flow-matching models in offline multi-objective optimization. It jointly optimizes a particle population using Wasserstein matching to an evolving objective-space proxy and directional transport toward the Pareto front. The paper provides an idealized convergence result and evaluates the method on 24 tasks using HV, GD, IGD, and Wasserstein distance.

Critical strengths
  1. I like the explicit distribution-level formulation. Matching the objective-space distribution provides a concrete way to control mass allocation alongside progress toward the front. Table 3 supports this motivation: directional transport alone can improve GD while worsening Wasserstein distance, whereas combining the two terms gives a better balance.

  2. I find the empirical evaluation convincing overall. The paper covers 24 tasks, five seeds, relevant forward and generative baselines, and complementary metrics beyond HV. The strong GD and Wasserstein rankings across all three benchmark families, together with the component ablations and runtime analysis, support the method’s practical value.

Critical weaknesses
  1. I see a substantial gap between the convergence theorem and the implemented algorithm. Theorem 1 assumes exact matching and access to the true Pareto-front projection, while the implementation uses learned surrogates, a fixed CHIM direction, and an evolving proxy. The remaining distributional gap is also not controlled. I therefore view the theorem as useful motivation rather than a guarantee for the practical method.

  2. I would like clearer evidence for the contribution of the flow prior itself. The ablations isolate the two guidance terms, but do not compare against directly optimizing the same transport objective from offline designs without flow dynamics. Since Appendix A.1.2 reports similar performance with zero regularization, this comparison would help establish how much the generative component adds.

Questions for authors
  1. Under what conditions is the fixed CHIM direction a reliable approximation to the projection direction? Do the observed matching and surrogate errors help explain cases where performance deteriorates?

  2. How does ParetoTransport compare with direct particle optimization using the same loss and proxy updates, initialized from offline designs under a comparable optimization budget?

  3. How are the Pareto-front reference points sampled for Wasserstein evaluation? Are the rankings stable to changes in reference-set density, given that uniform weights over reference points need not imply uniform mass along the continuous front?

2 comments

We thank the reviewer for their positive assessment. We are glad the reviewer likes the explicit distribution-level formulation and finds the empirical evaluation convincing, with strong GD and Wasserstein rankings supporting the method’s practical value. We address each weakness and question below.

Gap between Theorem 1 and implementation Theorem 1 indeed motivates our conceptual contribution and underlying assumption: the algorithm should perform well in cases where the offline front is sufficiently representative, i.e., where W2(μˉ,μ∗)W_2(\bar\mu, \mu^\ast) is small.

Contribution of the flow prior and Q2 Please note that λ=0\lambda = 0 (now clarified in Appendix A.1.3) does not remove the flow prior: the inner optimization is initialized at the posterior-mean estimate, and its solution is mapped back onto the sampling path of the pre-trained flow. However, we agree that a comparison with direct particle optimization is valuable and ran new experiments (Table 8 in Appendix A.1.1) on all tasks with the same setting, but without flow dynamics. With the flow prior, the surrogate error εf\varepsilon_f is lower and all metrics are better on DTLZ and RE, suggesting that the generative component keeps designs in regions where the surrogates are more reliable.

Table 8. Changes Δ relative to the offline front D(best), averaged over all tasks.

PriorGD (↓)IGD (↓)W2W_2 (↓)HV (↑)εf\varepsilon_f
ZDT✓-0.121-0.150-0.096+0.4650.087
✗-0.085-0.160-0.087+0.4500.085
DTLZ✓-0.179-0.142-0.174+0.3370.059
✗-0.120-0.126-0.090+0.3070.065
RE✓-0.026-0.056-0.035+0.4100.039
✗+0.059-0.010+0.119+0.2050.055

Q1 a) Fixed CHIM direction

The fixed CHIM normal nn is reliable when the projection directions of the offline-front points point along it on average, i.e. when cos⁡(Δˉ,n)≈1\cos(\bar\Delta, n) \approx 1 for the mean projection displacement Δˉ:=Ey∼μ0[ΠM∗(y)−y]\bar\Delta := \mathbb E_{y\sim\mu_0}[\Pi_{\mathcal M^\ast}(y) - y]. The average is the relevant criterion because LnL_n scores only the displacement of the population mean, while the Wasserstein term matches the individual points to the proxy. We added a new evaluation (Table 9 in Appendix A.1.1) of this alignment on all tasks. It is ≥0.99\ge 0.99 on 12 of 24 tasks and ≥0.78\ge 0.78 on 20, including concave and disconnected fronts.

Q1 b) Matching and surrogate errors We added a new evaluation (Table 10 in Appendix A.1.1) that decomposes the error on five representative tasks where ParetoTransport is worse than the offline front in W2W_2. On these tasks, the cause is either a large residual W2(μˉ,μ∗)W_2(\bar\mu, \mu^\ast) or a large surrogate error εf\varepsilon_f. On RE31, RE32 and RE33, the average matching error remains small, i.e. the particles reach the proxy, and GD(best) is small, but the residual W2(μˉ,μ∗)W_2(\bar\mu, \mu^\ast) is large and the final W2W_2 stays at or above its level. On the highly multimodal ZDT4 and DTLZ3, the surrogate error εf\varepsilon_f is large.

Table 10.

Task GD(best) IGD(best) Offline W2W_2 Residual Final W2W_2 Matching εf\varepsilon_f
ZDT40.0510.0680.4330.4270.550±0.0120.186±0.0490.242±0.026
DTLZ30.3410.0830.3410.0020.421±0.0060.088±0.0180.203±0.018
RE310.0170.4020.4520.4480.466±0.0720.011±0.0020.173±0.104
RE320.0510.0640.2960.2820.298±0.0180.021±0.0020.084±0.075
RE330.0070.0580.1720.1720.224±0.0500.057±0.0230.004±0.002

Q3 Reference-set density

We use the reference sets provided in the benchmark repository [1] and assign equal weight to each point. These sets are in general not uniform along the front. We therefore ran new evaluations (Table 11 in Appendix A.1.2) of W2W_2 with reference sets of the same size resampled uniformly by arc length (two objectives) or surface area (three objectives). The top rank is stable under this change, with ParetoTransport first in average rank in all three benchmark families.

Table 11. Avg. rank under W2W_2, reported → resampled.

MethodZDTDTLZRE
MultiHead-Vanilla4.000→3.8005.714→5.4295.667→5.917
MultipleModels-Vanilla6.400→6.2004.286→4.2865.500→5.375
ParetoFlow9.400→9.4008.714→8.5006.500→6.000
PGD7.800→7.8008.571→8.5718.208→7.667
ParetoTransport (ours)3.200→3.0001.571→1.5714.792→4.292

[1] K. Xue et al., Offline Multi-Objective Optimization. ICML 2024.

I thank the authors for their detailed responses and additional experiment for addressing my concerns. I raised my score from 3 to 4 to support the acceptance of this paper.

Comments on the paper

2 comments

We thank all reviewers for their time, effort, and constructive feedback. Below, we provide detailed, point-by-point responses to each individual review.

We have also uploaded a version with all changes highlighted in the supplementary material.

We thank all reviewers for the engaged discussion. Reviewer a8fq raised their score to 4 (clear acceptance), and Reviewer eKTU maintains their positive assessment (3). Reviewer 21HY stated that our discussion improved their evaluation by half a point (2.5) and recognized the work as a convincing and principled framework.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.