- 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 . While this is quite understandable, I would be interested in seeing what happens if the initial distribution 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
I thank the authors for detailed responses and well designed additional experiments. I will keep my positive assessment of this paper.
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+ and IGD+ [1]: d+(a,z)≤∥a−z∥ and thus GD+≤GD≤W2 and IGD+≤IGD≤W2. We have added new evaluations (Tables 47-53 in Appendix A.7) of GD+ and 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 W2 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.
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 W2 require. Following the reviewer’s suggestion, we have added the 13 MO-NAS tasks with m≤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 W2 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.
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), 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∗), 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<0.5. 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.