acceptodds
Under review as a conference paper at ICLR 2027

Meta Pareto Flow Learning with Ordinary Differential Equation Residual Supervision

Abstract

Pareto set learning (PSL) amortizes multiobjective optimization by learning a continuous mapping from objective preferences to solutions. However, existing PSL methods mainly model task-specific Pareto sets, whose terminal solution manifolds can be geometrically misaligned across related tasks, making shared transferable structure difficult to capture from these mappings alone. We propose Pareto Flow Learning (PFL), which formulates each preference-conditioned scalarized problem as a continuous-time Pareto gradient flow and trains a neural model to approximate the resulting flows through ordinary differential equation (ODE) residual supervision. The learned model represents optimization dynamics in addition to Pareto-optimal solutions, providing complementary process-level information for transfer. We further extend PFL to meta-learning by learning shared structure across source-task Pareto flows and adapting the resulting meta Pareto flow model (mPFM) to unseen targets. Integrating mPFM with multiobjective Bayesian optimization yields mPFM-MOBO for search under limited evaluation budgets. Experiments on synthetic benchmarks and engineering design problems show improved early-stage convergence and competitive or better final performance on most target tasks, including under distribution shift. Ablation studies further indicate that Pareto-flow ODE supervision provides useful transferable information beyond terminal Pareto-set mappings alone.

Primary area:
optimization
Keywords:
Multiobjective Optimization, Pareto Set Learning, Bayesian Optimization

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_m1X3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Pareto Flow Learning, a Pareto set learning technique that formulates each scalarized sub-problem as a continuous-time Pareto gradient flow, learning to map a preference vector to a near-optimal solution. The authors propose to learn this mapping via meta-learning, where the flow is conditioned on a two-part task embedding. The first part of the task embedding is computed from the training set, while the second part includes a lower-dimensional target-task-specific embedding. During meta-training, the target-specific embedding is learned via bilevel optimization. Lastly, the authors propose to use the learned flow-mapping in Multi-Objective Bayesian Optimization (MOBO), where the pretrained flow-mapping is used to optimize the acquisition function, moving the optimization from the decision space to the conditioned solution manifold. Furthermore, the authors also show how the task-specific encoding can be fine-tuned during active learning.

Overall the paper is well motivated and tackles an interesting problem. The empirical evaluation show that the i) proposed approach improves both the convergence speed and asymptotic performance over existing MOBO methods, ii) the Pareto flow formulation provides a better learning signal than using only terminal loss, and iii) the task embedding module has significant impact on the final performance. However, the empirical evaluations are limited mostly synthetic problems, with low dimensional decision spaces, limiting the generalizability of the claims.

Critical strengths
  • The paper introduces a Pareto gradient flow-based objective, which is extended to meta-learning setting and provide technical details on how it can be computed efficiently. The authors showcase empirically that the flow-based objective provides a more effective learning target than directly minimizing the terminal error.
  • The authors divide the learned task embedding into pretrained task embedding edataτe_{data}^{\tau} and target task-specific lower-dimensional embedding eoptτe_{opt}^{\tau}. This allows one to finetune only a small part of the model parameters during active learning, improving the efficiency. Moreover, the authors recognize this as a crucial component for improving the performance of their proposed method.
Critical weaknesses

My main concerns are related to evaluations:

  1. The authors consider only synthetic functions and 4 engineering problems, all of which have low-dimensional decision spaces. One of the claims in the Bayesian Optimization section is that the acquisition function optimization is easier in the conditioned manifold. While this may be true, I would like to see the method evaluated in more complicated problems with higher number of decision variables, such as trajectory planning [1] or vehicle design problem [2] to further support this claim.
  2. The authors only consider hypervolume as their performance metric; While this metric is commonly used in the literature, it is known to be quite sensitive to the used reference point [3]. It would be good to consider other metrics such as IGD+ and R2 to support the claims, especially since the ground truth solution sets are available for the synthetic tasks. Moreover, the Pareto-Front visualizations show only the proposed method, making it impossible to determine how e.g. diverse the generated solutions are compared to the baselines.

[1] Wang, Z., Gehring, C., Kohli, P. & Jegelka, S.. Batched Large-scale Bayesian Optimization in High-dimensional Spaces. In Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics, 2018

[2] T. Kohira, H. Kemmotsu, O. Akira, and T. Tatsukawa. Proposal of benchmark problem based on real-world car structure design optimization. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, 2018

[3] H. Ishibuchi, R. Imada, Y. Setoguchi and Y. Nojima, How to Specify a Reference Point in Hypervolume Calculation for Fair Performance Comparison, in Evolutionary Computation, vol. 26, no. 3, pp. 411-440, 2018

Questions for authors
  • The meta-loss function (Eq. 11) consists of many terms. How were the coefficients λIC,λODE,λTer,λReg\lambda_{\mathrm{IC}}, \lambda_{\mathrm{ODE}}, \lambda_{\mathrm{Ter}}, \lambda_{\mathrm{Reg}} selected for the loss, and how sensitive is the loss is to these coefficients?
  • The proposed approach seems to underperform in mDTLZ1 and mDTLZ$^{-1}$, where the Pareto Front is a simplex. Is there any intuition why this is the case?
  • Table 2 recognizes the target task adaption eoptτe^{\tau}_{\mathrm{opt}} as an important component of the mPFM-MOBO. Could you clarify if any of the baselines used a similar mechanism? If not, could a similar mechanism be applied to existing transfer PSL methods such as ExTrEMO or invTrEMO?

See also weaknesses section.

7 comments
AuthorsOpenReview ↗

Response (1/5)

Response to W1

To further evaluate our method in higher-dimensional decision spaces, we have conducted additional experiments on a 60-dimensional multiobjective Rover trajectory planning problem, extended from Wang et al. (2018) based on their source code https://github.com/zi-w/Ensemble-Bayesian-Optimization. The problem optimizes 30 free B-spline control points with two conflicting objectives: path length and obstacle risk. We further parameterize the obstacle configurations to construct related source tasks and consider four target settings, including two in-distribution (i.e., Rover-1 and Rover-3) and two out-of-distribution tasks (i.e., Rover-2 and Rover-4). We compare mPFM-MOBO with the five existing baselines over 20 independent runs, using the same budget of 100 evaluations (20 initial + 80 sequential evaluations).

(1) Hypervolume

SettingmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EI
Rover-11.2061 ± 0.00430.1458 ± 0.05790.3811 ± 0.04130.7615 ± 0.28440.4018 ± 0.08520.4380 ± 0.0812
Rover-21.1956 ± 0.01770.1290 ± 0.04760.3846 ± 0.05010.7577 ± 0.23670.3806 ± 0.08400.4458 ± 0.0823
Rover-31.2043 ± 0.00520.1815 ± 0.06780.4702 ± 0.05580.8024 ± 0.13710.4966 ± 0.06740.5151 ± 0.0958
Rover-41.1440 ± 0.01580.2693 ± 0.08640.5403 ± 0.04480.9581 ± 0.17890.5473 ± 0.05890.5893 ± 0.0735

(2) IGD+

SettingmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EI
Rover-10.0056 ± 0.00243.4973 ± 0.34122.2941 ± 0.18071.1483 ± 0.87292.1677 ± 0.36531.9174 ± 0.3134
Rover-20.0079 ± 0.00343.5749 ± 0.31792.4050 ± 0.25421.2005 ± 0.78452.2380 ± 0.31551.9815 ± 0.3719
Rover-30.0219 ± 0.00383.8556 ± 0.34782.4407 ± 0.22721.1491 ± 0.48452.1828 ± 0.35082.1176 ± 0.3751
Rover-40.0108 ± 0.01893.6881 ± 0.64322.2606 ± 0.21320.7916 ± 0.62352.1472 ± 0.28982.0076 ± 0.3045

(3) R2

SettingmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EI
Rover-10.0387 ± 0.00160.5691 ± 0.05300.3914 ± 0.02470.2144 ± 0.12700.3820 ± 0.05150.3607 ± 0.0457
Rover-20.0520 ± 0.00810.5902 ± 0.04140.3979 ± 0.02960.2211 ± 0.10980.4034 ± 0.05170.3653 ± 0.0449
Rover-30.0206 ± 0.00270.5213 ± 0.05650.3262 ± 0.02960.1727 ± 0.05920.3132 ± 0.03540.3040 ± 0.0501
Rover-40.0670 ± 0.00640.4670 ± 0.06040.3077 ± 0.02260.1281 ± 0.07110.3048 ± 0.02940.2846 ± 0.0363

As shown in the tables, mPFM-MOBO consistently achieves the best performance across all four target settings and all three metrics. Specifically, it achieves HV values of 1.144–1.206, substantially outperforming the strongest baseline, invTrEMO (0.758–0.958). Consistent improvements are also observed in IGD+ and R2. These results show that, for 60 dimensional problems, mPFM-MOBO also has a good performance.

Moreover, we clarify that our claim concerns improving Bayesian optimization's sample process, rather than reducing the computational difficulty of acquisition function optimization itself. By combining structured candidate generation with acquisition-based selection, the proposed method increases the likelihood of identifying high-quality solutions under limited evaluation budgets. We have updated the above results in the new uploaded manuscripts, please see Appendix I for details. We also have cited both recommended references [1, 2] in the revised appendix.

AuthorsOpenReview ↗

Response (2/5)

Response to W2 (Part 1)

To provide a more comprehensive evaluation beyond hypervolume, we have additionally reported IGD+ and R2 on the synthetic benchmarks, covering in-distribution and out-of-distribution settings. We also have updated the results and the recommended reference in Appendix J of the updated manuscript.

(1) IGD+

In-distribution Settings

ProblemmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EGO
mDTLZ10.0297±0.00510.3470±0.11670.2019±0.04500.1873±0.12050.2522±0.08310.6020±0.1599
mDTLZ20.0391±0.01110.1562±0.01360.1708±0.01650.1342±0.04690.1162±0.02410.2512±0.0346
mDTLZ30.0411±0.00700.2424±0.04740.2172±0.02650.1857±0.09200.1824±0.02750.4158±0.0789
mDTLZ40.0355±0.00280.1502±0.01440.1811±0.02030.1521±0.03390.1340±0.02000.2516±0.0381
mDTLZ1⁻¹0.0223±0.00130.2023±0.03540.1661±0.02650.0769±0.03900.1276±0.04080.3988±0.1697
mDTLZ2⁻¹0.0285±0.00180.0815±0.00610.0879±0.00680.0694±0.01350.0457±0.00510.2232±0.0426
mDTLZ3⁻¹0.0313±0.00300.1437±0.06640.1157±0.01120.0654±0.00750.0605±0.01230.2414±0.0466
mDTLZ4⁻¹0.0388±0.00410.0823±0.01110.1237±0.01610.0764±0.01870.0337±0.00170.2723±0.0449

Out-of-distribution Settings

ProblemmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EGO
mDTLZ10.0630±0.01400.7950±0.11460.4029±0.12800.7581±0.49140.7180±0.13360.9751±0.1491
mDTLZ20.0436±0.01130.1331±0.01550.1959±0.02730.1923±0.07320.1150±0.02340.2403±0.0427
mDTLZ30.0505±0.01160.3063±0.04990.2377±0.02670.2888±0.12940.2406±0.04310.4331±0.0490
mDTLZ40.0371±0.00370.1402±0.01420.1981±0.02710.2721±0.06880.1202±0.01350.2448±0.0280
mDTLZ1⁻¹0.0960±0.15780.6818±0.12310.3562±0.07910.3288±0.35240.4560±0.15350.7276±0.1393
mDTLZ2⁻¹0.0331±0.00490.1006±0.00740.1099±0.00910.0999±0.01940.0578±0.00670.2167±0.0325
mDTLZ3⁻¹0.0323±0.00260.2084±0.03360.1413±0.01840.1109±0.01430.0958±0.03040.3164±0.0359
mDTLZ4⁻¹0.0346±0.00360.0781±0.00860.1235±0.01320.0914±0.01110.0365±0.00540.2481±0.0388

(2) R2

In-distribution Settings

ProblemmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EGO
mDTLZ10.0379±0.00120.0856±0.01270.0817±0.01020.0635±0.01970.0712±0.01190.1263±0.0214
mDTLZ20.0966±0.00130.1125±0.00250.1135±0.00210.1073±0.00630.1159±0.01060.1343±0.0215
mDTLZ30.0969±0.00100.1251±0.00680.1196±0.00420.1128±0.01100.1262±0.01190.1527±0.0272
mDTLZ40.0953±0.00040.1101±0.00200.1132±0.00200.1090±0.00370.1078±0.00190.1214±0.0050
mDTLZ1⁻¹0.1043±0.00180.2128±0.02250.1977±0.01520.1286±0.01870.1694±0.03050.2749±0.0408
mDTLZ2⁻¹0.1250±0.00060.1339±0.00130.1451±0.00260.1371±0.00440.1286±0.00080.1976±0.0145
mDTLZ3⁻¹0.1257±0.00110.1611±0.03650.1529±0.00350.1354±0.00220.1334±0.00290.2051±0.0139
mDTLZ4⁻¹0.1299±0.00130.1364±0.00270.1548±0.00360.1391±0.00690.1281±0.00060.2049±0.0141
AuthorsOpenReview ↗

Response (3/5)

Response to W2 (Part 2)

Out-of-distribution Settings

ProblemmPFM-MOBODirHV-EGOExTrEMOInvTrEMOPSL-MOBOPSL-DirHV-EGO
mDTLZ10.0460±0.00320.1406±0.01150.1069±0.01140.1378±0.06510.1328±0.01850.1579±0.0115
mDTLZ20.0981±0.00150.1090±0.00250.1187±0.00420.1127±0.00820.1211±0.01160.1397±0.0236
mDTLZ30.0989±0.00160.1282±0.00540.1229±0.00560.1239±0.01560.1389±0.01290.1576±0.0259
mDTLZ40.0955±0.00050.1094±0.00180.1160±0.00300.1221±0.00750.1060±0.00170.1216±0.0052
mDTLZ1⁻¹0.1319±0.04710.3107±0.00300.2853±0.02550.2040±0.08970.2916±0.03390.3121±0.0037
mDTLZ2⁻¹0.1263±0.00160.1417±0.00270.1531±0.00390.1472±0.00630.1334±0.00220.1930±0.0099
mDTLZ3⁻¹0.1257±0.00090.1844±0.01610.1645±0.00550.1501±0.00480.1467±0.01140.2295±0.0137
mDTLZ4⁻¹0.1285±0.00140.1350±0.00180.1554±0.00420.1446±0.00370.1285±0.00170.1969±0.0116

The additional results are largely consistent with our HV-based findings. In particular, mPFM-MOBO achieves the best IGD+ and R2 values on most synthetic tasks, demonstrating its effectiveness in approximating the Pareto front under limited evaluation budgets.

Regarding the Pareto front visualizations, we clarify that the original figures illustrate the performance of the learned Pareto set models. Among the baselines, PSL-MOBO, PSL-DirHV-EGO, and invTrEMO can generate such Pareto set model through their learning component, whereas the remaining methods do not provide an equivalent preference-conditioned solution generator. We have therefore added visual comparisons with these three methods in Appendix E, Figure 8 - 11, allowing a more direct assessment of the convergence and diversity of the learned Pareto sets. The visualizations further show that mPFM tends to provide more accurate approximations to the Pareto fronts on most benchmark problems, demonstrating the effectiveness of the proposed method.

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

The paper proposes Pareto Flow Learning (PFL). Each preference-conditioned augmented-Tchebycheff subproblem is treated as a gradient-flow ODE. A network u^(w,t)\hat{u}(\mathbf{w}, t) is trained with an initial-condition loss and an ODE residual, PINN-style. It is extended to a meta-model (mPFM) with a shared backbone and head, a set-encoder embedding edatae_{\text{data}}, and a task-specific eopte_{\text{opt}}. The mPFM is plugged into GP-based multiobjective Bayesian optimization as a candidate generator (mPFM-MOBO). Experiments on mDTLZ, mDTLZ$^{-1}$, and four engineering problems show faster early HV convergence than PSL-MOBO, DirHV-EGO, ExTrEMO, and invTrEMO.

Critical strengths
  • The paper is clearly written and the pipeline is easy to follow.
  • Using ODE residual supervision to learn optimization trajectories, and framing transfer as learning shared structure across Pareto flows, is an interesting idea.
  • The evaluation is reasonably broad: in- and out-of-distribution targets, engineering problems, batch acquisition, and trajectory visualizations.
Critical weaknesses
  • Scope is narrower than the story. PFL (Section 3) and meta-PFL (Section 4) are presented as general frameworks, but they are only evaluated inside black-box MOBO with ≤100\leq 100 evaluations. PFL alone is never tested. Missing settings are white-box MOO (e.g., LibMOON-style benchmarks against Pareto set learning and hypernetwork baselines) and multi-task learning. The claim that ODE supervision "improves learning and transfer of optimization dynamics" is not supported outside this one pipeline.
  • The central concepts are never defined. "Shared structure," "shared dynamical representation," "geometrically misaligned terminal manifolds," and "task-specific manifold" are used informally. No assumption specifies how tasks relate, and no evidence shows that flows are more alignable than Pareto sets.
  • The paper has no same-architecture PFL/PSL trained only on target data, optimizing eopte_{\text{opt}} against the GP LCB, with no source pretraining. Without it, gains cannot be attributed to meta-learning rather than to test-time optimization of eopte_{\text{opt}} against the LCB. Table 2 supports the concern: with eopte_{\text{opt}} frozen, HV collapses (e.g., 0.000.00 on mDTLZ1/3 OOD), so the performance comes mostly from the 500-step LCB adaptation.
  • The ablation w/ vs. w/o ODE should be discussed deeply in the main manuscript.
  • Limited novelty in the meta component. The conditioning design (shared backbone, task embeddings, residual modulation Δhτ=rγ([edata,eopt])\Delta h^{\tau} = r_{\gamma}\left([e_{\text{data}}, e_{\text{opt}}]\right)) closely resembles shared-hypernetwork/adapter approaches such as Hyperformer++ (Mahabadi et al., 2021). The novelty rests mainly on ODE supervision. PHN-style Pareto hypernetworks (Navon et al., 2021) are also missing from the related work.
  • Little analysis of when and why it works. There is no controlled study varying task distance (e.g., δ1,δ2\delta_1, \delta_2 gradations) or showing negative transfer.

References

Mahabadi, R. K., Ruder, S., Dehghani, M., & Henderson, J. (2021). Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 565–576.

Navon, A., Shamsian, A., Chechik, G., & Fetaya, E. (2021). Learning the Pareto front with hypernetworks. In International Conference on Learning Representations (ICLR).

Questions for authors

Please see weaknesses.

5 comments
AuthorsOpenReview ↗

Response (1/3)

Response to W1

We thank the reviewer for this suggestion. To investigate the applicability of pretrained meta Pareto flow representations to white-box PSL, we additionally develop a flow-guided PSL extension (mPFM-White) that combines a pretrained mPFM with a trainable Pareto set model. Specifically, the pretrained mPFM backbone is frozen, with only its task-specific embedding adapted, while the Pareto set model is trained using a scalarized multiobjective loss. An annealing mechanism progressively shifts solution generation from the pretrained flow to the Pareto set model. We compare this extension against four LibMOON baselines (Agg-Tche, Agg-SoftTche, EPO, and PMGDA) on eight mDTLZ/mDTLZ$^{-1}$ problem families under in-distribution and out-of-distribution settings. All methods use the same 100 training iterations and 50 batch size of preferences per iteration, with HV results evaluated on 1,000 preference-conditioned predictions over five random seeds.

ProblemSettingmPFM-WhiteAgg-TcheAgg-SoftTcheEPOPMGDA
DTLZ1In-Distribution1.0488±0.01141.0364±0.01580.9217±0.03500.8010±0.03981.0025±0.0094
DTLZ1Out-of-Distribution1.0516±0.01331.0413±0.00331.0182±0.04400.8962±0.03211.0003±0.0051
mDTLZ2In-Distribution0.7565±0.00320.7603±0.00240.5159±0.11090.7196±0.01470.6943±0.0043
mDTLZ2Out-of-Distribution0.7560±0.00680.7502±0.00960.4606±0.00100.7025±0.02250.6813±0.0019
mDTLZ3In-Distribution0.7568±0.00110.7527±0.00760.4603±0.00040.5942±0.10780.6826±0.0038
mDTLZ3Out-of-Distribution0.7526±0.01040.7430±0.01720.3245±0.16600.5840±0.07170.6735±0.0062
mDTLZ4In-Distribution0.7572±0.00210.7615±0.00080.6825±0.11060.7322±0.00490.7129±0.0044
mDTLZ4Out-of-Distribution0.7580±0.00280.7618±0.00190.6297±0.13860.7277±0.00420.7066±0.0042
mDTLZ1$^{-1}$In-Distribution0.3073±0.00390.2076±0.03940.0846±0.00030.0210±0.01650.0025±0.0003
mDTLZ1$^{-1}$Out-of-Distribution0.3087±0.00210.2647±0.02030.0868±0.00060.0487±0.02210.0024±0.0012
mDTLZ2$^{-1}$In-Distribution0.7583±0.00060.7578±0.00140.4267±0.00720.7318±0.00210.0018±0.0005
mDTLZ2$^{-1}$Out-of-Distribution0.7565±0.00090.7548±0.00210.3989±0.00800.5102±0.05360.0017±0.0002
mDTLZ3$^{-1}$In-Distribution0.7561±0.00300.7510±0.00280.3432±0.00410.3903±0.02480.0023±0.0001
mDTLZ3$^{-1}$Out-of-Distribution0.7503±0.00320.7438±0.01180.3443±0.00320.3514±0.02980.0020±0.0006
mDTLZ4$^{-1}$In-Distribution0.7595±0.00070.7596±0.00040.4381±0.00440.7084±0.01480.0016±0.0004
mDTLZ4$^{-1}$Out-of-Distribution0.7578±0.00140.7577±0.00100.4304±0.01400.7401±0.00380.0017±0.0002

The results show that mPFM-White achieves the highest mean HV on 12 of 16 settings and ties at the reported precision on one additional setting, while maintaining competitive performance on most remaining tasks. These findings suggest that pretrained Pareto flow representations can provide useful guidance for white-box Pareto set learning.

Nevertheless, we acknowledge that extending mPFL to high-dimensional decision spaces, particularly neural-network-based multi-task learning, remains challenging because accurately learning preference-conditioned optimization trajectories can entail substantial computational and modeling complexity. We have therefore clarified this limitation and narrowed the scope of our main claims to multiobjective Bayesian optimization, while presenting the white-box PSL results as an exploratory extension rather than a comprehensive solution to general white-box multiobjective learning. We also have modified our title and Section I to limit the scope mainly in multiobjective Bayesian optimization. All of the modified parts have been highlighted via blue.

AuthorsOpenReview ↗

Response (2/3)

Response to W2

We thank the reviewer for pointing out the need for clearer definitions. We have clarified these concepts in Section 4 (page 4) of the updated manuscript under “Transferable Information in Meta Pareto Flow Learning.” Specifically, for each task τ\tau, we define the task-specific solution geometry at time tt as Pτ(t)=[uτ(w,t)∣w∈W]\mathcal P_\tau(t)=[\boldsymbol u_\tau(\boldsymbol w,t)\mid\boldsymbol w\in\mathcal W], where uτ\boldsymbol u_\tau denotes the preference-conditioned gradient flow. At t=tmax⁡t=t_{\max}, this set approximates the task-specific Pareto set. We refer to discrepancies between the terminal sets of different tasks as terminal geometric misalignment. Our key motivation is that different terminal Pareto sets do not necessarily imply entirely different optimization dynamics. Related tasks may exhibit similar intermediate solution geometries or descent directions over a nontrivial interval I⊂(t0,tmax⁡)\mathcal I\subset(t_0,t_{\max}), even when their terminal solutions are substantially different. We refer to these common patterns as shared dynamical structure, which motivates learning a shared dynamical representation over preferences and optimization time, rather than transferring only terminal Pareto solutions. Thus, our task-relatedness assumption is that related tasks may share transferable optimization dynamics despite differences in their terminal solutions.

Response to W3

We thank the reviewer for raising this important concern. To isolate the contribution of meta-learning, we additionally introduce a non-pretrained PFL baseline that uses the same network architecture and target-task adaptation procedure, including the 500-step optimization of eopte_{\mathrm{opt}} against the GP-LCB objective, but without source-task meta-pretraining.

ProblemSettingMeta PretrainedNon-Pretrain
mDTLZ1In-Distribution1.0126 ± 0.02470.0000 ± 0.0000
mDTLZ1Out-of-Distribution0.8556 ± 0.06320.0000 ± 0.0000
mDTLZ2In-Distribution0.6727 ± 0.02740.0390 ± 0.0381
mDTLZ2Out-of-Distribution0.6551 ± 0.02670.0129 ± 0.0198
mDTLZ3In-Distribution0.6663 ± 0.01850.0025 ± 0.0056
mDTLZ3Out-of-Distribution0.6366 ± 0.02970.0000 ± 0.0000
mDTLZ4In-Distribution0.6935 ± 0.00730.0491 ± 0.0207
mDTLZ4Out-of-Distribution0.6878 ± 0.00960.0103 ± 0.0236

As shown in the table, meta-pretraining substantially improves performance across all eight settings, whereas the non-pretrained model achieves very low HV despite undergoing the same 500-step LCB adaptation. These results demonstrate that test-time optimization of eopte_{\mathrm{opt}} alone cannot explain the observed performance gains. Rather, meta-pretraining provides a transferable Pareto flow representation that enables effective target-task adaptation. Together with the frozen-embedding ablation in Table 2, these findings highlight the complementary roles of meta-pretraining and target adaptation, i.e., pretraining provides transferable optimization knowledge, while LCB-based adaptation tailors this knowledge to the target task. We have added this comparison to the revised manuscript (Appendix K, page 29).

Response to W4

We thank the reviewer for this suggestion. We have expanded the ODE supervision ablation and moved it into the main manuscript (Section 6.3, page 9), with additional analysis of why and when ODE supervision benefits meta-learning. First, unlike terminal Pareto-set supervision, which constrains only the final preference-to-solution mapping, ODE supervision additionally constrains intermediate optimization trajectories and descent directions, providing richer learning signals across related tasks. To further investigate when these additional constraints are beneficial, we introduce three source-task groups with increasing diversity in benchmark-family composition. Under Group 1, where the source tasks have relatively similar Pareto-set structures, Meta-PSL remains competitive, suggesting that terminal supervision can already capture sufficient shared information. Under Groups 2 and 3, which include more heterogeneous problem families, mPFM substantially outperforms Meta-PSL on both target tasks. These results reveal that the benefit of ODE supervision is particularly pronounced when heterogeneous source tasks make terminal Pareto-set supervision less effective. This provides a more detailed explanation of the role of optimization dynamics in meta Pareto set learning, beyond the performance improvements observed in the original ablation.

AuthorsOpenReview ↗

Response (3/3)

Response to W5

We agree that the conditioning architecture, including the shared backbone, task embeddings, and residual modulation, builds upon established designs rather than introducing a fundamentally new neural architecture. In fact, our main contribution lies in investigating whether ODE residual supervision can improve meta Pareto set learning in multiobjective Bayesian optimization, rather than proposing a novel hypernetwork architecture. We have also added discussions and citations of Hyperformer++ (Mahabadi et al., 2021) and Pareto HyperNetworks (Navon et al., 2021) to better distinguish our contribution from existing conditional and preference-based neural representations. Please refer to Remark 1 on page 5 and Appendix A.2 on page 17, for details.

Response to W6

We thank the reviewer for this suggestion. To investigate when knowledge transfer becomes less effective, we conduct two additional experiments examining target-task distance and source-task heterogeneity, respectively.

(1) Sensitivity to task distance. We systematically vary the target-task parameters (δ1,δ2)(\delta_1,\delta_2) across five configurations, ranging from the original in-distribution setting to increasingly distant target tasks. The resulting HV performance on mDTLZ1–4 is reported below.

δ1\delta_1δ2\delta_2mDTLZ1mDTLZ2mDTLZ3mDTLZ4
0.70.251.0126±0.02470.6727±0.02740.6663±0.01850.6935±0.0073
0.60.350.8556±0.06320.6551±0.02670.6366±0.02970.6878±0.0096
0.50.350.7674±0.18640.6359±0.02160.6194±0.03520.6856±0.0184
0.40.40.6059±0.00550.5931±0.03000.5652±0.05650.6748±0.0256
0.30.50.9266±0.30890.5377±0.03790.5240±0.09940.6330±0.0833

The results show that HV consistently decreases on mDTLZ2–4 as the target-task parameters move further from the original setting, indicating reduced transfer effectiveness under greater source-target discrepancies. However, mDTLZ1 exhibits non-monotonic behavior, suggesting that transfer performance also depends on the underlying optimization landscape rather than task distance alone.

(2) Negative transfer We further investigate whether jointly learning from structurally different source tasks can introduce transfer interference. Specifically, we compare the original mPFM jointly trained on eight benchmark families with family-specific models trained separately on mDTLZ1 and mDTLZ1⁻¹, while varying task parameters within each family.

ProblemSettingMeta-trainingHV ↑IGD+ ↓R2 ↓
mDTLZ1IDGeneral1.0126 ± 0.02470.0297 ± 0.00510.0379 ± 0.0012
mDTLZ1IDFamily-specific1.0189 ± 0.03040.0294 ± 0.00740.0368 ± 0.0013
mDTLZ1OODGeneral0.8556 ± 0.06320.0630 ± 0.01400.0460 ± 0.0032
mDTLZ1OODFamily-specific0.9704 ± 0.07860.0528 ± 0.02130.0387 ± 0.0030
mDTLZ1⁻¹IDGeneral0.2624 ± 0.00460.0223 ± 0.00130.1043 ± 0.0018
mDTLZ1⁻¹IDFamily-specific0.2749 ± 0.02300.0203 ± 0.00830.1002 ± 0.0036
mDTLZ1⁻¹OODGeneral0.1780 ± 0.07240.0960 ± 0.15780.1319 ± 0.0471
mDTLZ1⁻¹OODFamily-specific0.2262 ± 0.06520.0771 ± 0.15870.1171 ± 0.0497

As shown above, family-specific meta-training generally improves performance, particularly on OOD tasks, where HV increases by approximately 13.4% on mDTLZ1 and 27.1% on mDTLZ1⁻¹ compared with joint eight-family training. These results provide evidence of potential transfer interference arising from heterogeneous source tasks, suggesting that large differences in Pareto-set geometry and optimization landscapes can bring additional challenge shared meta-learning.

We have added the corresponding analyses to the revised manuscript (Appendix L).

Reviewer_y9yrOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines ODE-supervised Pareto flow learning with meta-learning for sample-efficient MOBO.

Critical strengths

The approach is conceptually appealing, with ablations supporting trajectory supervision.

Critical weaknesses

Controlled benchmarks limit practical relevance, while offline costs and shared dynamics remain insufficiently examined.

Questions for authors

Is source supervision practical to obtain, and do gains justify its acquisition and training costs?

1 comment

Response to W and Q

Thanks for the comment from the reviewers. We mainly response from the following for points.

Regarding the practicality of source supervision and offline costs, our method follows an offline meta-learning and online transfer optimization paradigm. During offline training, mPFM learns transferable optimization knowledge from related source tasks through ODE residual supervision, without requiring precomputed Pareto-optimal trajectories or ground-truth Pareto sets. In practical applications, particularly industrial optimization, historical data from previous operations or simulations can be used to construct surrogate models that provide the objective evaluations and gradient estimates required for training, thereby reducing the need for additional expensive evaluations of the original systems. During online optimization, the pretrained mPFM is adapted to unseen target tasks using limited objective evaluations, transferring previously learned knowledge to accelerate optimization. Importantly, offline training is performed only once, and the learned representation can be reused across multiple target tasks. Thus, its computational cost can be amortized across subsequent optimization tasks, while the primary benefit lies in improving the sample efficiency of expensive multiobjective optimization.

Regarding shared dynamics, we have provided additional experimental evidence to examine the role of optimization dynamics in cross-task transfer in Section 6.3 of the updated manuscript. Specifically, we have expanded the architecture-matched Meta-PSL ablation and introduced experiments involving different source-task distributions. The results show that ODE supervision becomes particularly beneficial when more heterogeneous benchmark families are included, suggesting that intermediate optimization dynamics provide useful transferable information beyond terminal Pareto-set mappings. Together with the task-distance sensitivity analysis, these experiments offer further empirical insights into when such dynamical information benefits transfer and where its limitations arise.

Regarding the practical relevance of controlled benchmarks, our evaluation already includes several engineering design problems alongside synthetic benchmarks, which enable systematic assessment of cross-task transfer under controlled task variations and distribution shifts, please see Appendix F, page 22. In the revised manuscript, we further extend the evaluation to a 60-dimensional multiobjective Rover trajectory planning problem, involving conflicting path-length and obstacle-risk objectives under different environment configurations. Please see Appendix I, page 26. These experiments combine controlled evaluation with application-oriented optimization scenarios, providing meaningful evidence of the effectiveness and potential practical applicability of our approach. While a gap remains between benchmark-based evaluation and real-world deployment, our primary objective is to investigate transferable Pareto flow learning for sample-efficient multiobjective Bayesian optimization, rather than deployment in a specific industrial application.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Major Reviewer Concerns and Our Responses

Dear Area Chair and Reviewers,

Thank you for considering our submission and the rebuttal, and also thanks for the discussion with the reviewers. Our paper proposes mPFM-MOBO, which uses ODE supervision to learn preference-conditioned optimization trajectories from related source tasks and guide candidate generation on new tasks under limited evaluation budgets. The reviewers’ main concerns were the scope of the evaluation, the respective roles of meta-pretraining and ODE supervision, and the conditions and offline requirements for effective transfer. In response, we added higher-dimensional and white-box experiments, matched controls and ablations, and analyses of sensitivity and negative transfer. Reviewer wUfE indicated that most of their concerns had been resolved and raised their rating from 3 to 4. Reviewer m1X3 stated that their concerns had been addressed and updated their rating to 3.

  • Evaluation scope and metrics. We added a 60-dimensional Rover trajectory-planning benchmark. Under the same budget of 100 evaluations, mPFM-MOBO achieved the best mean results among the compared methods on hypervolume, IGD+, and R2 across all four target settings. We also added IGD+ and R2 on the synthetic benchmarks, Pareto-front visualizations alongside three baselines, and an exploratory white-box Pareto set learning study against LibMOON baselines.

  • Meta-pretraining, ODE supervision, and target adaptation. A same-architecture control with identical target adaptation achieved substantially lower hypervolume without source pretraining. The expanded ODE ablation showed that trajectory supervision was especially useful as the source-task families became more heterogeneous. We also clarified how our target adaptation differs from the transfer-GP adaptation used by ExTrEMO and invTrEMO, and clarified the paper’s relationship to prior hypernetwork work.

  • Sensitivity and limits of transfer. We tested sensitivity to the main loss weights and to source–target discrepancy. We also compared joint meta-training with family-specific training on mDTLZ1 and its inverted variant. Family-specific training improved performance, particularly on out-of-distribution targets, providing evidence of interference between source families in those settings.

  • Offline supervision. In response to y9yr, we clarified that training uses source-task objective and gradient information without requiring precomputed Pareto-optimal trajectories or ground-truth Pareto sets. The pretrained model can be reused across target tasks, allowing its training cost to be amortized across applications.

Following wUfE’s remaining suggestion, we will sharpen the mathematical definitions in Section 4, especially what it means for tasks to share dynamical structure and how that relates to differences between their terminal Pareto sets.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.