acceptodds
Under review as a conference paper at ICLR 2027

MOGF/D: Decomposition and Cooperation in Gradient Flow for Pareto Set Learning

Abstract

Gradient-based multi-objective optimization aims to construct a finite set of solutions representing diverse Pareto trade-offs, but individual gradient descent does not by itself determine how these solutions should organize as a population. Existing methods can generate multiple trade-offs through decomposition, preference adaptation, or particle interaction, but they do not explicitly separate per-subproblem descent from population-level cooperation. To address this, we propose a normal-tangent cooperation mechanism that makes population interaction gradient-orthogonal to scalar descent, which preserves first-order decrease and, after a finite formation phase, leads to an exact Lyapunov identity and conditional convergence to Pareto-critical accumulation points. Experiments on synthetic MOO benchmarks and real learning tasks show that the proposed method remains competitive on simple continuous fronts and achieves substantially larger set-quality gains on disconnected and multimodal Pareto problems.

Primary area:
optimization
Keywords:
Pareto Set Learning, Gradient Flow, MOEA/D

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_CNtgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

I was assigned the review of this paper one day before the deadline.

Multiobjective optimization is a key problem in many applications. This paper aims to construct a finite set of solutions and Pareto tradeoffs by proposing an cooperation mechanism. Experiments on synthetic benchmarks should demonstrate the usefullness of the approach.

Critical strengths

The approach with the normal tangent is clearly derived and potentially can even be called elegant. The paper is not too dense and easy to read.

Critical weaknesses

I am not sure how the theorem / propositions directly links to what is done in the experiments and they just to seem to be restatements e.g. Proposition 2.1? And Theorem 3.1 or Lemma C3 This seems to be straightforwad from the definitions or restatements and not necessairly what is actually needed to cover the experiments? What would actually be very strong and maybe be an independent publication would be better coverage, faster converge but the above seems to be fairly straightforward results and not necessairly be applicable to how the experiments are implemented.

Several Ablations of each model choice seem to be missing.

It is unclear how munch tuning was done for the number of hyperparameters a full ablation / sensitivity analysis and it seems that code is not to be released or at least nowhere mentioend?

Questions for authors

see above

4 comments
AuthorsOpenReview ↗

Response by the authors (1/3)

Thank you for noting that our normal–tangent derivation is "clearly derived and potentially can even be called elegant" and easy to read. Below we address (1/3) W1, (2/3) W2 (ablations), and (3/3) W3 (tuning, sensitivity, and code).

W1: Constructive Theory, Coverage, and Convergence Speed

W1: I am not sure how the theorem / propositions directly links to what is done in the experiments and they just to seem to be restatements e.g. Proposition 2.1? And Theorem 3.1 or Lemma C3 This seems to be straightforwad from the definitions or restatements and not necessairly what is actually needed to cover the experiments? What would actually be very strong and maybe be an independent publication would be better coverage, faster converge but the above seems to be fairly straightforward results and not necessairly be applicable to how the experiments are implemented.

Lemma C.3, Proposition 2.1, and Theorem 3.1 are not restatements: they form a constructive chain governing our two-stage algorithm and explaining both wider coverage and faster convergence.

Constructive chain and implementation. Lemma C.3 proves gi=∇ϕi(xi)g_i=\nabla\phi_i(x_i) is a strictly positive softmax mixture of objective gradients, so scalar stationarity (ni=0n_i=0) gives exact Pareto criticality in O(md)O(md) closed form without per-particle QPs (MGDA, ParetoMTL). Naive repulsion violates box bounds and pushes converged points off the front; Proposition 2.1 projects cooperation onto free coordinates LX(xi)L_{\mathcal{X}}(x_i) with ε2\varepsilon^2-damping and cap ∥ti∥2≤ρ∥ni∥2\lVert t_i\rVert_2\le\rho\lVert n_i\rVert_2, ensuring zero cooperative work (⟨gi,ni+ti⟩=−∥ni∥22\langle g_i,n_i+t_i\rangle=-\lVert n_i\rVert_2^2) and equilibrium invariance (vi=0  ⟺  ni=0v_i=0\iff n_i=0) for all ρ≥0\rho\ge 0. Since curvature LL perturbs tangency over finite steps, Theorem 3.1 pairs alignment σ=(1+ρ2)−1/2\sigma=(1+\rho^2)^{-1/2} with Armijo bound κ=min⁡(σamin⁡M,2β(1−c)σ2L)>0\kappa=\min\bigl(\frac{\sigma a_{\min}}{M},\frac{2\beta(1-c)\sigma^2}{L}\bigr)>0 to yield O(H−1)O(H^{-1}) residual decay over H=T−KH=T-K frozen steps and local linear contraction (1−η)k−K0(1-\eta)^{k-K_0} on PL basins. Practically, formation (k<Kk<K) adapts preferences and scales to allocate basins; freezing ϕˉi=ϕiK\bar{\phi}_i=\phi_i^K at k=Kk=K activates Proposition 2.1 and Theorem 3.1 under 17-trial Armijo on analytic tasks and minibatch Adam or Adam–Armijo on neural models.

Better coverage. Without claiming a global nonconvex coverage theorem, Proposition 2.1 lifts the coverage–descent ceiling of unprojected updates (ni+cin_i+c_i). On curved or density-biased Pareto sets, repulsive and gap forces (Proposition B.2) carry uphill terms (⟨gi,ci⟩>0\langle g_i,c_i\rangle>0): weak weights (ρ≪1\rho\ll 1) leave wide front gaps, while strong weights (ρ≥1\rho\ge 1) flip updates uphill and stall line search. Enforcing ⟨gi,ti⟩=0\langle g_i,t_i\rangle=0 unlocks strong level-set transport (ρ≥1\rho\ge 1) while keeping subproblems synchronized across preference rays.

Faster convergence. Conserving full normal dissipation −∥ni∥22-\lVert n_i\rVert_2^2 in O(md)O(md) closed form eliminates cooperative drag, uphill stalls, and QP overhead, accelerating steps, epochs, evaluations, and wall-clock time (Table 1).

Table 1. Link from normal–tangent geometry to coverage, speed, and updaters (12 seeds).

Dimension & SettingMOGF/D (Orthogonal)Unprojected / BaselineMechanistic Takeaway
Coverage (F9, K=3501)Fill h 0.040; IGD 0.0091h 0.114 (none) / 0.203 (unproj.)-65% to -80% gap; #1 14/21
Step speed (F9, K=3501)0.0287 in 99 steps; 1.84M tot0.0285 (1,500, none); 2.13M unproj.15x fewer steps vs scalar-only
Epoch speed (MSLR, 10−410^{-4})HV 8.07 (ep 60), 8.26 (90)8.01 ep 150 (none); 8.13 unproj.2.5x faster epochs; less stall
Wall-clock & equal-time F725.3s avg; F7 (18.8s) 0.0294PMTL 215.1s; WFR (18.8s) 2.0388.5x faster vs PMTL; 0% uphill
Stage & neural updatersF7/F9 at T: 0.025/0.0091At K: 0.163/0.169; Armijo 0% riseFrozen stage refines across updaters
AuthorsOpenReview ↗

Response by the authors (2/3)

W2: Systematic Ablations Across Synthetic and Neural Model Choices

W2: Several Ablations of each model choice seem to be missing.

Thank you for raising this point. Holding budgets, initializations, and readouts fixed, we ablated every core design choice across our continuous and neural Pareto-set learning pipelines.

Synthetic pipeline and scaling ablations. Table 2a isolates five synthetic choices: (i) cooperative direction, comparing orthogonal projection against equal-norm unprojected (Matched), raw unprojected, and scalar-only (None) updates from random starts and shared frozen states at KK; (ii) two-stage scheduling, comparing pre-freeze states at KK against frozen refinement at TT; (iii) individual formation mechanisms, removing Gaussian repulsion, archive-gap guidance, preference adaptation, or coordinate proposals one by one; (iv) disabling coordinate proposals across all 17 transfer benchmarks, showing that orthogonal projection independently improves set quality whenever nonlinear linkage (LZ09) or density bias (DTLZ4) couples lateral spreading with descent, while remaining neutral on separable problems; and (v) problem scaling across d=10,30,100d=10,30,100 and M=2,3,5M=2,3,5.

Neural optimizer, population, and backbone ablations. Table 2b isolates four neural choices: (i) updaters crossed with orientation (Adam, Adam–Armijo, Direct Armijo) on Adult and six-objective MSLR-WEB10K; (ii) single-batch monotonicity (Adam–Armijo minus Adam), confirming that batch non-increase gives no test gain whereas preserving scalar normal progress consistently improves held-out trade-offs; (iii) population size N=6,9,12,18N=6,9,12,18; and (iv) model scale, widening MultiMNIST by 32.9x in parameters and evaluating six-model ResNet18 populations on NYUv2 and Cityscapes.

Table 2a. Synthetic ablations across orientation, stage schedule, components, and scale.

Ablated ChoiceBenchmark & ProtocolMOGF/D (Orthogonal)Ablated / Control Variants
Direction (full/K)LZ09-F7 (12 seeds, IGD)Full 0.0294; tail 0.0294Matched 0.0400/0.0354; None 0.0469/0.0412
Suite direction17 tasks (12 seeds, no prop.)7/17 dual wins (pH<0.05p_{\text{H}}<0.05)Wins on DTLZ4, F2–F5, F7, F9 (8 neutral)
Stage (K -> T)LZ09-F7 / F9 (12 seeds, IGD)At T: 0.0251 / 0.0091At K: 0.1630 (-84.6%) / 0.1693 (-94.6%)
ComponentsZDT3 / F7 (6 seeds, IGD)Full 0.0097 / 0.0320No gap 0.0102/0.0336; fix w 0.0459; no prop. 0.1752
Scale (d, M)d=10,30,100; M=2,3,5F7 0.025 / 0.101 / 0.269Matched 0.035 / 0.118 / 0.284; DTLZ2 to M=5

Table 2b. Neural ablations across optimizers, orientations, population size, and scale.

Ablated ChoiceTask & MetricMOGF/D (Orthogonal)Ablated / Baseline Variants
3 Updaters ×\times dir.Adult (12 seeds, Traj. HV)0.6744 / 0.6743 / 0.6748Matched 0.6733 / 0.6733 / 0.6738 (pH≤0.018p_{\text{H}}\le 0.018)
6-obj dir. (Adam/Armijo)MSLR (12 seeds, 10−410^{-4} HV)Val-Sel 8.606 / 8.534Matched 8.323/8.155 (pH≤0.024p_{\text{H}}\le 0.024); None 8.194/8.169
Batch monotonicityAdam–Armijo minus AdamAdult +0.24/-0.09/-0.06 (10−310^{-3})MSLR -0.07/+0.01/-0.03 (10−410^{-4}); all pH=1.00p_{\text{H}}=1.00
Population size NMSLR (N=6, 9, 12, 18)HV 5.73->8.15->8.81->9.60Monotonic 6D test HV (10−410^{-4}) gain (500 ep)
Model scaleMultiMNIST, NYUv2, CityscapesCNN 7.780; 0.198 / 0.546Matched 7.773; PMTL 0.180/0.537; WFR 0.183/0.542
AuthorsOpenReview ↗

Response by the authors (3/3)

W3: Hyperparameter Tuning, Sensitivity, Zero-Retuning Transfer, and Code Release

W3: It is unclear how munch tuning was done for the number of hyperparameters a full ablation / sensitivity analysis and it seems that code is not to be released or at least nowhere mentioend?

Thank you for raising these reproducibility points. Below we detail our tuning scope, sensitivity sweeps, zero-retuning transfer across 17 benchmarks, and code release.

Tuning protocol and validation-only selection. Formation parameters were chosen once via staged Bayesian optimization (Optuna) on development runs—locking prior fields at each stage—and held fixed thereafter (Table 3b). On all neural tasks, screening, selection, and checkpoint readout maximized validation hypervolume on optimization seeds disjoint from evaluation, with the test split disabled throughout tuning. Early synthetic searches overlapped part of the initial six-seed cohort; the revised paper discloses this explicitly and re-verifies all mechanism claims on fresh 12-seed cohorts.

Sensitivity grid and 17-benchmark zero-retuning transfer. Sweeping the cooperation multiplier eight-fold from shared frozen states on LZ09-F9 (Table 3a) shows that orthogonal projection monotonically improves both front proximity (IGD) and worst-case uncovered gap (Fill h) while keeping rejected steps below 0.5%, whereas unprojected cooperation works only when heavily damped before collapsing into uphill rejections. Learning-rate and cap controls on Adult and MSLR-WEB10K likewise show stable test gains without step-size fragility. Most decisively, freezing four family configurations and transferring them with zero per-problem retuning and zero coordinate proposals across 17 canonical benchmarks (2,040 runs) ranks MOGF/D #1 overall, ruling out per-task tuning (Table 3b).

Table 3a. Cooperation-strength sensitivity on LZ09-F9 (12 shared-K seeds, cap 4λ4\lambda).

λ\lambda (Cap)MOGF/D IGD ↓\downarrowUnproj. IGD ↓\downarrowFill h (Orth / Unproj)Zero (Orth / Unproj)
00 (00)0.0285 ± 0.00420.0285 ± 0.00420.114 / 0.1140.03% / 0.03%
1/81/8 (0.50.5)0.0229 ± 0.00480.0133 ± 0.00320.104 / 0.0700.15% / 0.25%
1/41/4 (1.01.0)0.0177 ± 0.00400.0154 ± 0.00620.086 / 0.0730.15% / 0.05%
1/21/2 (2.02.0)0.0119 ± 0.00250.0965 ± 0.02870.062 / 0.2010.17% / 90.25%
11 (4.04.0)0.0091 ± 0.00210.1000 ± 0.03010.040 / 0.2030.47% / 91.79%

Table 3b. Tuning protocol and 17-benchmark zero-retuning transfer (2,040 runs).

Protocol / SuiteTuning & Config RuleStanding (8 Methods)Sensitivity / Control
Synthetic OptunaStaged search (rate, gap, width)Disjoint confirm seedsLocked for 12-seed controls
Neural OptunaVal-HV screen + confirm seedsTest disabled in tuningStable across LR & ρ\rho
ZDT4,6 & DTLZ1–6Frozen ZDT1,2 & DTLZ7 configs#1 4/8, Top-2 6/8Zero retuning; wins DTLZ4
LZ09-F1–F9 (9 tasks)Frozen F9 config (no proposals)#1 6/9, Top-2 9/9Wins 6/9 controls (F7)
All 17 tasks4 frozen configs total#1 10/17, Top-2 15/17AvgRank 1.76/8 (WFR 2.59)

Code availability and open-source release. We apologize for omitting an explicit code-release notice initially. An anonymized reproducibility package (Supplementary Material zip on OpenReview) is already uploaded—containing executable synthetic and neural code for MOGF/D and baselines, fixed configurations, Optuna logs, per-seed outputs, and verification scripts alongside Algorithms 1–3 in the revised paper—and we commit to public open-source release upon publication (to be stated explicitly in the Reproducibility Statement).

We hope our response has addressed your concerns. If we have missed anything, please let us know.

AuthorsOpenReview ↗

Official Response by Authors

Dear Reviewer CNtg,

Thank you very much for taking the time to carefully review our work and provide your constructive comments. Your feedback has been invaluable in helping us improve our work.

In response to your comments, we have included the following content:

  • We showed how Lemma C.3, Proposition 2.1, and Theorem 3.1 form a constructive chain governing our two-stage method and explaining its wider coverage (-65% to -80% gap on LZ09-F9) and faster convergence (8.5x faster than PMTL on LZ09-F7) (Section 3, Appendix D.2; Response (1/3)).
  • We added systematic fixed-budget ablations isolating cooperative orientation (7/17 dual wins without coordinate proposals), two-stage freezing (K -> T), scaling in (d, M), neural optimizers, and larger backbones (Appendices F.8–F.16; Response (2/3)).
  • We detailed the validation-only tuning protocol with fresh 12-seed verification, confirmed 17-benchmark zero-retuning transfer (avg. rank 1.76/8), and uploaded an anonymized reproducibility package with public code release committed (Appendices B.1, F.7; Response (3/3)).
  • Besides, as suggested by Reviewers DotQ and qTsf, we proved finite-backtracking satisfaction of Eq. (16) (Appendix D.2), added Algorithms 1–3, and expanded theoretical comparisons against MOO-SVGD, PNG, and MGDA/D (Appendix A).

May we kindly ask if our response has addressed the issues you raised? If you still have any further questions or concerns, we would be more than happy to continue the discussion.

Thank you again for your time and effort.

Best regards,

Authors of Submission 41554

Reviewer_sSk6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MOGF/D, which separates scalar descent from population-level cooperation in multi-objective optimization through a normal–tangent gradient flow. The cooperative direction is constructed to be orthogonal to the scalar gradient, thereby preserving first-order descent. The method adopts a two-stage optimization strategy and provides a Lyapunov analysis and conditional convergence guarantees for the frozen dynamics. Although the experiments show promising results on synthetic multi-objective benchmarks and real learning tasks, the independent contribution of the core mechanism and the consistency between the theoretical analysis and the practical algorithm remain insufficiently validated.

Critical strengths
  • The normal–tangent decomposition provides a clear geometric interpretation of population-level cooperation, supported by a Lyapunov identity that characterizes its descent-preserving property.

  • The method achieves promising results on problems with complex Pareto geometries, including ZDT3, DTLZ7, and LZ09-F7, with quantitative and visualization results demonstrating its potential for improving Pareto-set coverage.

Critical weaknesses
  1. There is a substantial gap between the theoretical guarantees and the actual optimization procedure.

    The Lyapunov identity and convergence analysis apply only to the frozen normal–tangent dynamics, whereas the practical algorithm includes a formation stage that is not covered by the same theoretical guarantees. More importantly, this stage accounts for most of the computational budget in several key experiments. For example, ZDT3 enters the frozen stage only after 4,901 of 5,001 updates, while DTLZ7 freezes after 921 of 1,001 updates. Consequently, the theoretically analyzed dynamics account for only a small fraction of the optimization process, making it difficult to directly explain the final empirical performance.

    Furthermore, Theorem 3.1 relies on an additional finite-step nondegeneracy condition (Equation 16), while the real learning experiments use minibatch gradients and Adam rather than the projected Armijo recursion considered in the analysis.

  2. The independent contribution of the core normal–tangent cooperation mechanism is insufficiently validated.

    The main methodological novelty is projecting population-level cooperation onto a direction orthogonal to the scalar gradient. However, the current ablations primarily remove interaction, gap guidance, and preference adaptation, without directly comparing the same interaction signal with and without the normal–tangent projection.

    Moreover, the observed benefits of cooperation are relatively limited in some cases. On ZDT3, removing interaction increases IGD only from 0.00968 to 0.00975, while HV remains nearly unchanged. On LZ09-F7, removing interaction increases IGD from 0.03202 to 0.03828. In addition, none of the paired component comparisons in Table 9 reaches statistical significance at the 0.05 level after Holm correction. The current experiments do not convincingly establish that the proposed orthogonal cooperation mechanism is the key factor behind the performance improvements.

  3. The main experimental gains are affected by additional search mechanisms and inconsistent experimental settings.

    LZ09-F7 is one of the benchmarks where the paper reports its most substantial improvements. However, this experiment additionally employs random single-coordinate perturbation proposals for basin exploration, which are not part of the theoretical normal–tangent gradient-flow framework.

    Table 8 in Appendix F.4 shows that removing coordinate proposals significantly worsens IGD from 0.03202 to 0.17524, whereas removing interaction increases IGD only to 0.03828. This suggests that the major improvements on LZ09-F7 may be driven primarily by the additional basin-exploration mechanism rather than the proposed cooperation mechanism itself.

    Particle-WFR uses 30 decision variables on LZ09-F7, while MOGF/D and the other methods use only 10, meaning that this key comparison is not conducted under identical problem dimensionality. Different methods also employ different definitions of optimization updates and potentially different computational costs. Therefore, the current experimental evidence does not sufficiently isolate the performance gains attributable to the core algorithmic design of MOGF/D.

Questions for authors

See weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer sSk6

Thank you for your careful review and thoughtful comments. We address your concerns point by point below.

W1: The theoretically analyzed phase is short on some tasks, and finite-step assumptions and Adam complicate its link to empirical performance.

Response: Formation prepares the subproblem allocation and accessible search regions. Frozen cooperation then organizes trajectories while preserving fixed scalar descent. Table 19 now quantifies this division: F7 returned IGD improves from 0.163001 at KK to 0.025094, and F9 from 0.169278 to 0.009107. The common-state controls in Tables 17 and 20 isolate the additional direction effect after formation is shared. The short ZDT3/DTLZ7 tails are also reported. Proposition D.2 and Appendices F.11/F.14 connect the ideal construction to finite-step behavior and actual learning updates, respectively. Section 3 now explains these complementary roles directly.

W2: The original component removals and nonsignificant tests do not establish the independent contribution of projection.

Response: Tables 17–18 supply the missing direct intervention. F7 improves by 26.4%, with 11/12 paired wins and Holm p=0.041992p=0.041992 in the planned two-test family. Its shared-state continuation improves by 16.9% over norm-matched ordinary cooperation. F9 improves by 86.3%, with Holm p=0.005371p=0.005371 in the separate 11-setting family. Table 20 and Figure 13 further explain the mechanism across the full strength grid, including the effective low-strength ordinary regime. The original component-removal statistics remain separately reported, so the new evidence directly answers the projection question.

W3: F7's coordinate proposals confound attribution, and its WFR comparison differs in dimension and per-update cost.

Response: The common-state controls inherit the same coordinate-exploration outcome and disable proposals thereafter, establishing a subsequent cooperative increment. We also added a corrected d=10d=10, N=50N=50 comparison at matched measured optimization time in Appendix F.12, Table 22. MOGF/D's 5001 updates take 18.773 ± 0.159 s, giving IGD 0.029413 ± 0.008717. WFR completes 9056–9238 updates within the same per-seed core time and gives 2.038418 ± 0.512074. MOGF/D wins all 12 pairs. The table specifies the retained configurations and timing scope, including archive maintenance. Section 4.1 now directs the F7 comparison to this dimension- and cost-controlled result.

I would like to thank the authors for their hard work and the detailed comment.

The additional experiments and clarifications have completely resolved my concerns regarding W2 and W3.

Regarding W1, I find the newly added Table 19 to be highly valuable, as it directly and effectively demonstrates the necessity and impact of the two optimization stages. I strongly recommend incorporating this experiment into the main text in future versions of the paper. Furthermore, I suggest appropriately expanding Section 3 to provide a deeper explanation of the respective roles of these two stages and their complementary relationship, ideally accompanied by corresponding theoretical derivations to make the mechanism even more rigorous.

Overall, I am very satisfied with the authors' response and the improvements made to the manuscript. I believe my current rating is reasonable and justified, and I stand by my positive assessment.

AuthorsOpenReview ↗

Response to Reviewer sSk6

Thank you for your encouraging feedback and helpful suggestion.

W1 (follow-up): Further explain the roles and complementarity of the two optimization stages, with supporting theoretical derivations, and incorporate Table 19 into the main text.

Response: We are pleased that W2 and W3 have been resolved. The two stages separate adaptive population organization from cooperative refinement under fixed scalar objectives; their complementarity can be made explicit as follows.

Let ϕik\phi_i^k denote the scalar potential with the preferences and normalization at iteration kk. During formation, its change decomposes exactly as

ϕik+1(xik+1)−ϕik(xik)=ϕik(xik+1)−ϕik(xik)+Dik,\phi_i^{k+1}(x_i^{k+1})-\phi_i^k(x_i^k)=\phi_i^k(x_i^{k+1})-\phi_i^k(x_i^k)+D_i^k,

and

:=\phi_i^{k+1}(x_i^{k+1})-\phi_i^k(x_i^{k+1}).$$ The first term measures movement under the current scalar objective; $D_i^k$ captures changes in that objective as preferences and normalization adapt. Formation therefore allows the population to adjust both its allocation of trade-offs and its location. At the finite horizon $K$, we fix $\bar\phi_i=\phi_i^K$, so $D_i^k=0$ for $k\ge K$. The population and cooperative signals continue to evolve, while the orthogonal construction gives $\langle g_i,n_i+t_i\rangle=-\|n_i\|^2$, where $g_i=\nabla\bar\phi_i(x_i)$ and $n_i$ is the projected scalar-descent direction. Cooperation can thus redirect trajectories while preserving instantaneous scalar dissipation, as formalized by Proposition 2.1. A finite-budget consequence follows directly from Theorem 3.1. Under its assumptions, including the accepted-step conditions (15)–(16), we have $$\bar\phi_i(x_i^{k+1}) \le \bar\phi_i(x_i^k)-c\kappa\|n_i^k\|^2+\epsilon_k.$$ Summing over $H$ frozen updates and using the lower bound $\underline\phi_i=\inf_{x\in\mathcal X}\bar\phi_i(x)$ yields $$\min_{K\le k<K+H}\|n_i^k\|^2 \le \frac{\bar\phi_i(x_i^K)-\underline\phi_i+ \sum_{k=K}^{K+H-1}\epsilon_k}{c\kappa H}.$$ Here $c$ is the Armijo parameter, $\kappa>0$ is the displacement constant, and the nonnegative acceptance errors are summable. Hence the smallest squared projected-stationarity residual is bounded by $O(H^{-1})$. **Formation determines the frozen subproblems and the complete state from which refinement begins; the frozen stage then provides a quantitative descent-based refinement guarantee.** Table 19 complements this analysis by measuring population-level set quality at the transition and after the remaining optimization. We will incorporate Table 19 into the main text and integrate this explanation into Section 3. We expect to complete the revision within the next few days, after consolidating the remaining reviewers’ feedback and arranging the new results within the main-text page limit. Thank you again for helping us make this connection clearer.
Reviewer_qTsfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MOGF/D, a gradient-flow method for Pareto set learning. It decomposes the descent direction of each scalar subproblem and the population-level cooperative direction into normal and tangential components, ensuring that the cooperative motion is orthogonal to the scalar gradient and therefore preserves first-order descent. The paper further analyzes the Lyapunov identity and conditional convergence of the frozen phase, while introducing a finite formation phase to improve the population distribution.

Critical strengths

1- MOGF/D separates individual descent from population-level cooperation and preserves the first-order decrease of each scalar subproblem through orthogonal tangential cooperation. It further establishes a Lyapunov identity and provides a conditional convergence analysis, enhancing the interpretability of the method.

2- The paper evaluates the method on synthetic benchmarks, fairness classification, and learning-to-rank tasks. The results indicate that it is effective for disconnected fronts and multimodal Pareto sets.

Critical weaknesses

1- Theorem 3.1 directly assumes the finite-step nondegeneracy condition, but the paper does not prove or explain that the projection, backtracking, and step-size strategies used in practice satisfy this condition.

Questions for authors

1- Can the authors prove that the strategies used in the actual implementation satisfy the finite-step nondegeneracy condition in Eq. (16)? Alternatively, could they provide additional interpretability analyses or empirical experiments to assess this condition?

2- Compared with MOO-SVGD, PNG, MGDA/D, and other methods that project repulsive or population-level signals onto the gradient-orthogonal subspace, what are the rigorous mathematical novelties of MOGF/D in terms of its construction and convergence guarantees?

3- How were the hyperparameters for preference adaptation, archive guidance, interaction, and coordinate proposals selected during the formation stage?

4- The paper relies primarily on textual descriptions in the main paper and appendix, but it lacks complete pseudocode.

5- In Appendix F.1, the reference point for MSLR is written as 060^6. It would be clearer to denote it explicitly as the six-dimensional zero vector 0∈R6\mathbf{0}\in\mathbb{R}^6.

5 comments
AuthorsOpenReview ↗

Response by the authors (1/3)

Thank you for your constructive review and for recognizing MOGF/D's orthogonal cooperation, Lyapunov and conditional convergence analysis, and effectiveness on disconnected and multimodal fronts. Our self-contained response has three parts: (1/3) W1 & Q1 on Eq. (16) in practice, (2/3) Q2 on mathematical novelties, and (3/3) Q3–Q5 on hyperparameter selection, zero-retuning transfer, pseudocode, and notation.

W1 & Q1: Finite-Step Nondegeneracy Condition Eq. (16) in Practice

W1: 1- Theorem 3.1 directly assumes the finite-step nondegeneracy condition, but the paper does not prove or explain that the projection, backtracking, and step-size strategies used in practice satisfy this condition.

Q1: 1- Can the authors prove that the strategies used in the actual implementation satisfy the finite-step nondegeneracy condition in Eq. (16)? Alternatively, could they provide additional interpretability analyses or empirical experiments to assess this condition?

In exact arithmetic, our box-feasible tangential projection and geometric Armijo backtracking provably satisfy Eq. (16) with an explicit constant κ>0\kappa > 0. In practice, departures arise solely from single-precision rounding once solutions are already near Pareto stationarity, or from minibatch variance under Adam—never from cooperative interference with descent.

Exact-arithmetic bound and single-precision floor. On X=[l,u]d\mathcal{X}=[l,u]^d, masking outward gradients at active faces yields ni=ProjTX(xi)(−gi)n_i=Proj_{T_{\mathcal{X}}(x_i)}(-g_i), and projecting cooperation orthogonally onto free coordinates LX(xi)L_{\mathcal{X}}(x_i) under ∥ti∥2≤ρ∥ni∥2\lVert t_i\rVert_2\le\rho\lVert n_i\rVert_2 guarantees alignment σ=(1+ρ2)−1/2\sigma=(1+\rho^2)^{-1/2} between uk=vk/∥vk∥2u_k=v_k/\lVert v_k\rVert_2 and −gk-g_k. Along LL-smooth interior segments (0<∥gk∥2≤M0<\lVert g_k\rVert_2\le M, initial step α0,k≥amin⁡>0\alpha_{0,k}\ge a_{\min}>0, Armijo c,β∈(0,1)c,\beta\in(0,1)), every α≤2(1−c)σL∥gk∥2\alpha\le\frac{2(1-c)\sigma}{L}\lVert g_k\rVert_2 passes Armijo, giving −⟨gk,dk⟩≥κ∥gk∥22,κ=min⁡(σamin⁡M,  2β(1−c)σ2L)>0.-\langle g_k, d_k\rangle \ge \kappa\lVert g_k\rVert_2^2, \qquad \kappa = \min\left(\frac{\sigma a_{\min}}{M},\; \frac{2\beta(1-c)\sigma^2}{L}\right) > 0. Under 17 trials (Bmax⁡=16B_{\max}=16, α0,k≤amax⁡\alpha_{0,k}\le a_{\max}), this bound holds down to ∥gk∥2≥δ>0\lVert g_k\rVert_2\ge\delta>0 whenever amax⁡β16≤2(1−c)σδLa_{\max}\beta^{16}\le\frac{2(1-c)\sigma\delta}{L}. At tiny residuals in float32 (median ∥ni∥2=1.27×10−3\lVert n_i\rVert_2=1.27\times 10^{-3} on LZ09-F7), required decreases −c gi⊤di-c\,g_i^\top d_i fall below single-precision resolution, yielding benign near-stationary zero steps that vanish in float64—whereas unprojected cooperation injects genuine uphill directions throughout optimization.

Minibatch neural training. Standard Adam occasionally raises single-batch losses via momentum and sampling noise rather than input geometry; adding minibatch Armijo (Adam–Armijo, Direct Armijo) eliminates same-batch increases and confirms consistent test hypervolume gains across updaters (Table 1).

Table 1. Finite-step descent and acceptance across synthetic and neural tasks (12 seeds each).

Evaluation SettingUpdate RuleMOGF/D (Orthogonal)Unprojected / Reference
LZ09-F7 precision audit17-trial Armijo29.14% active zero (float32)99.96% pass in float64 (100% in 19)
Standard synthetic suites17-trial Armijo0.00% zero (ZDT2, ZDT3); 0.47–3.10% (F9, F2, DTLZ7)Active residual ∥ni∥2>10−6\lVert n_i\rVert_2 > 10^{-6}
17-benchmark transfer17-trial Armijo0.00% active uphill (17/17); 0–3.54% all-step zero (6/8 LZ09)89.7–99.7% uphill & 89.5–98.9% zero (8 LZ09)
Adult neural fairnessAdam & Armijo0.00% batch inc. (Armijo); wins 12/12 Traj. HV8.48% batch increases under Adam
MSLR 6-objective LTRAdam & Armijo0.00% inc., 11.48% zero; Test HV 8.61 / 8.5337.94% zero (Armijo); Test HV 8.32 / 8.15 (10−410^{-4})
AuthorsOpenReview ↗

Response by the authors (2/3)

Q2: Rigorous Mathematical Novelties vs. MOO-SVGD, PNG, and MGDA/D

Q2: 2- Compared with MOO-SVGD, PNG, MGDA/D, and other methods that project repulsive or population-level signals onto the gradient-orthogonal subspace, what are the rigorous mathematical novelties of MOGF/D in terms of its construction and convergence guarantees?

While MOO-SVGD, PNG, and MGDA/D combine multi-objective gradients with population interaction or constrained updates, MOGF/D enforces a distinct per-particle invariant: frozen cooperation preserves exact first-order scalar dissipation while allowing inter-objective trade-offs and strong steering (ρ≥1\rho \ge 1) without disturbing Pareto optima (Table 2a).

Table 2a. Mathematical construction, descent property, and behavior at Pareto optima.

MethodCooperative MechanismDescent & Trade-Off PropertyAt Pareto Optima (ni=0n_i=0)
MOGF/D (Frozen)Closed-form map on LXL_{\mathcal{X}} (cap ρ\rho)Zero work (gi⊤ti=0g_i^\top t_i=0); allows trade-offsVanishes (vi=0⇔ni=0v_i=0\Leftrightarrow n_i=0)
MOO-SVGD (NeurIPS'21)Kernel MGD + −α∇xjkij-\alpha\nabla_{x_j}k_{ij}Density RKHS orthogonality, not pointwiseRepulsion pushes xix_i off Pareto set
PNG (UAI'22)mm-loss QP (∇ℓj⊤u≥φt\nabla\ell_j^\top u\ge\varphi_t)Simultaneous mm-loss descent; no trade-offsDrops constraints (φt=−∞\varphi_t=-\infty)
MGDA/D (GECCO'25)Scalarized GD + neighbor stepNeighbor transfer to escape local optimaUpdates stagnated solutions

Zero cooperative work and trade-off freedom (∀ρ>0\forall \rho > 0). Projecting bib_i onto the two-sided feasible box subspace LX(xi)=TX(xi)∩[−TX(xi)]L_{\mathcal{X}}(x_i)=T_{\mathcal{X}}(x_i)\cap[-T_{\mathcal{X}}(x_i)] via Tε,xi(gi)bi=∥gi,L∥22bi,L−⟨gi,L,bi,L⟩gi,L∥gi,L∥22+ε2T_{\varepsilon,x_i}(g_i)b_i=\frac{\lVert g_{i,L}\rVert_2^2 b_{i,L}-\langle g_{i,L},b_{i,L}\rangle g_{i,L}}{\lVert g_{i,L}\rVert_2^2+\varepsilon^2} ensures orthogonality to both gig_i and NX(xi)N_{\mathcal{X}}(x_i) in O(d)O(d) closed form, giving ⟨gi,ni+ti⟩=−∥ni∥22\langle g_i,n_i+t_i\rangle=-\lVert n_i\rVert_2^2 (Proposition 2.1) for any ρ>0\rho>0. Unlike PNG—which enforces simultaneous decrease of all mm losses away from stationarity and drops those constraints near stationarity—MOGF/D lets individual objectives trade off freely along ϕi\phi_i's level sets without consuming normal descent.

Stationary-set invariance and strong steering (ρ≥1\rho \ge 1). Scaling the cap by ∥ni∥2\lVert n_i\rVert_2 (∥ti∥2≤ρ∥ni∥2\lVert t_i\rVert_2\le\rho\lVert n_i\rVert_2) preserves equilibria (ni+ti=0  ⟺  ni=0n_i+t_i=0\iff n_i=0) and alignment (1+ρ2)−1/2>0(1+\rho^2)^{-1/2}>0 even for ρ≥1\rho\ge 1, whereas unprojected ni+cin_i+c_i guarantees descent only for ρ<1\rho<1. In MOO-SVGD on f1=f2=12x2f_1=f_2=\tfrac{1}{2}x^2 with x1=0x_1=0 (Pareto-optimal), x2=1x_2=1, k(x,y)=exp⁡(−(x−y)2/2)k(x,y)=\exp(-(x-y)^2/2), and α=1\alpha=1, g⋆(x)=xg^\star(x)=x yields φ^(0)=e−1/2≠0\widehat{\varphi}(0)=e^{-1/2}\ne 0, pushing x1x_1 off stationarity, whereas MOGF/D enforces n1=0  ⟹  t1=0n_1=0\implies t_1=0.

Table 2b. LZ09-F9 cooperation-strength sweep from shared frozen states (12 seeds, cap 4λ4\lambda).

λ\lambda (Cap)MOGF/D IGD ↓\downarrowUnprojected IGD ↓\downarrowActive ZeroUnprojected Zero
00 (00)0.0285 ± 0.00420.0285 ± 0.00420.03%0.03%
1/81/8 (0.50.5)0.0229 ± 0.00480.0133 ± 0.00320.15%0.25%
1/41/4 (1.01.0)0.0177 ± 0.00400.0154 ± 0.00620.15%0.05%
1/21/2 (2.02.0)0.0119 ± 0.00250.0965 ± 0.02870.17%90.25%
11 (4.04.0)0.0091 ± 0.00210.1000 ± 0.03010.47%91.79%

Unprojected cooperation works only when heavily damped and stalls into uphill rejections once the effective cap exceeds unity, whereas orthogonal projection sustains strong lateral steering for superior front coverage. Even when unprojected inputs descend (LZ09-F7, ρ=0.5\rho=0.5), preserving exact normal dissipation still yields significant IGD gains from both random and shared frozen states.

AuthorsOpenReview ↗

Response by the authors (3/3)

Q3: Hyperparameter Selection for the Four Formation Mechanisms

Q3: 3- How were the hyperparameters for preference adaptation, archive guidance, interaction, and coordinate proposals selected during the formation stage?

Staged Bayesian optimization (Optuna) selected fixed formation parameters on development runs (Table 3a), using validation hypervolume on disjoint seeds with the test split disabled for neural tasks. Early synthetic searches overlapped part of the six-seed cohort, and all mechanism claims were re-verified on fresh 12-seed cohorts. Freezing four representative configurations for zero-retuning, zero-coordinate-proposal transfer across 17 canonical benchmarks (12 seeds each) confirms that MOGF/D does not rely on per-task tuning or coordinate perturbations (Table 3b).

Table 3a. Search spaces and selection rules for the four formation mechanisms.

MechanismSynthetic Search SpaceNeural Search SpaceSelection Rule
Preference adaptationRate [1e-3, 0.08], anchor [0.5, 12]Rate [0, 0.08], ratio [0.1, 0.95]Front/gap or validation HV
Archive guidanceWeight [0.02, 5], gap [3, 50]Weight [1e-3, 0.3], warm-up 1–10Front/gap or validation HV
Particle interactionStrength [0.1, 6], width [0.02, 0.3]Strength [5e-4, 0.3], width [0.05, 0.8]Fixed; test split disabled
Coordinate proposalsPeriod {1..20}, scale [0.12, 0.4]Disabled in neural tasksHistorical LZ09-F7 only

Table 3b. Zero-retuning, zero-coordinate-proposal transfer (2,040 runs; dual wins pHolm<0.05p_{\text{Holm}}<0.05).

Benchmark SuiteSourceExternal Rank (8 Methods)Orthogonal vs. Controls
ZDT4, ZDT6ZDT1, ZDT2#1/8 on both (2/2)Neutral on separable tasks
DTLZ1–DTLZ6DTLZ7#1 DTLZ1–2, #2 DTLZ3–4Dual win on DTLZ4
LZ09-F1–F9 (no proposals)LZ09-F9#1 on 6/9, Top-2 9/9Wins 6/9 (incl. LZ09-F7)
Overall (17 tasks)4 configs#1 10/17, Top-2 15/177/17 dual wins (Avg. 1.76)

Q4: Complete Implementation-Level Pseudocode

Q4: 4- The paper relies primarily on textual descriptions in the main paper and appendix, but it lacks complete pseudocode.

Full pseudocode (Algorithms 1–3) is in the revised paper and summarized below:

Algorithm: Two-Stage MOGF/D (Formation -> Frozen Normal-Tangent Cooperation)
Require: F, population N, budget T, horizons (E, N_0, W_0), K = max(E, N_0, W_0), rho, eps = 1e-4.
1. Initialize {x_i}_{i=1}^N, interior weights {w_i}, and nondominated archive A = empty.
2. For k = 0, ..., T - 1:
3.   Evaluate F(X), update A, and update (z, s) for k < N_0 (freeze (z*, s*) at N_0).
4.   Compute g_i = grad phi_i(x_i; w_i, z, s), n_i = Proj_{T_X(x_i)}(-g_i), subspace L_i.
5.   Project b_i = lambda_k b_i^int + lambda_cov G a_i b_i^cov onto L_i:
       t_tilde_i = (||g_{i,L}||^2 b_{i,L} - <g_{i,L}, b_{i,L}> g_{i,L}) / (||g_{i,L}||^2 + eps^2),
       t_i = min(1, rho ||n_i|| / max(||t_tilde_i||, 1e-12)) * t_tilde_i,   v_i = n_i + t_i.
6.   If k < E (Formation): adapt w_i (if k < W_0); step along v_i (+ guard / F7 proposal).
7.   Else (Refinement; frozen for k >= K):
       - Synthetic: 17-trial Armijo on ClipInset(x_i+alpha v_i/||v_i||) (g_i^T d_i<0, Eq. 15; else x_i).
       - Neural: minibatch Adam or Adam-Armijo (grad = -v_i); select best validation-HV set.
8. Return size-N nondominated/quantile set from A U X (synthetic) or best validation-HV set (ML).

Q5: Explicit Notation for the MSLR Reference Vector

Q5: 5- In Appendix F.1, the reference point for MSLR is written as 06{0}^{6}. It would be clearer to denote it explicitly as the six-dimensional zero vector 0∈R6\mathbf{0} \in {\mathbb{R}}^{6}.

Thank you; Appendix F.1 now writes this reference as 0∈R6\mathbf{0} \in \mathbb{R}^6.

We hope our response has addressed your concerns. If we have missed anything, please let us know.

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

This paper proposes MOGF/D, which uses preference-conditioned scalarization and population cooperation to construct a multi-objective trade-off solution set. The authors give the continuous-time Lyapunov identity and conditional discrete convergence proof for the frozen phase, and report experimental results on synthetic problems, fairness classification and multi-objective ranking tasks.

After rebuttal, I rise my rating to weak accept.

Critical strengths
  1. The paper clearly distinguishes between individual optimization and population cooperation, and gives the derivation of maintaining scalar potential decrease during the frozen phase.

  2. The experiments consider continuous fronts, disconnected fronts and problems with multiple local Pareto sets.

Critical weaknesses
  1. It is hard to judge whether the performance improvement mainly comes from orthogonal cooperation or from population formation and archive readout. In Table 8, the impact of removing the interaction or gap guidance is relatively limited, while the impact of removing the coordinate proposals for F7 is more obvious. This shows that population formation plays an important role in the results, but existing ablations do not directly compare ordinary cooperation and orthogonal cooperation. In addition, Appendix B explains that the final results will also be selected from visited archives, and the quality of the final solution set is also affected by the readout rule.

  2. The role of the descent-preserving cooperation mechanism in learning experiments is unclear. The guarantees of Section 3 are for specified updates after freezing, while learning experiments use minibatch gradients and Adam. In addition, the checkpoints selected in the fairness experiment are from the formation stage or its boundary and cannot be used to verify the role of the subsequent frozen phase.

  3. Test problems have limited coverage. Section 4.1 only selects ZDT1 to ZDT3, DTLZ7, and F2, F7, and F9 in LZ09. The authors organized these experiments by geometric complexity but did not fully explain why other applicable problems in the same benchmark suites were not included. This is not yet enough to judge its overall performance. The decision-space dimensions and number of objectives in Table 3 also lack systematic variation, making it difficult to see the adaptability of the method to problem scale and different Pareto-front characteristics. More complete benchmarking would help clarify the scope of advantages and limitations of the approach.

  4. The coverage of ML applications is also relatively limited. The three fairness datasets adopt similar classification settings. The other application is multi-objective ranking, and the models used are small fully connected networks. These results support the effectiveness of the method on the selected tasks. However, validation on different task types, model structures, and problem sizes remains insufficient.

Questions for authors
  1. Keeping the population formation, archiving and final readout methods unchanged, and only comparing independent STCH, ordinary cooperation and orthogonal cooperation, can the orthogonal mechanism still bring stable benefits?

  2. After using Adam, does the actual update retain the emphasized descent properties? Can the effect of this property on performance be illustrated through controlled experiments with objective changes during training and corresponding update methods?

  3. Are the variants of the proposed method and its cooperation mechanism still effective on more complete benchmarks and in larger-scale ML scenarios?

8 comments
AuthorsOpenReview ↗

Response to Reviewer DotQ

Thank you for your careful review and thoughtful comments. We address your concerns point by point below.

W1 & Q1: Do the gains come from orthogonal cooperation itself, rather than formation, coordinate exploration, or archive readout? A direct comparison with ordinary cooperation and independent STCH is needed.

Response: The new controls isolate cooperative orientation. With the same formation, archive, readout, and budget rules, F7's 12 fresh paired seeds give IGD 0.029413 versus 0.039975 for orthogonal and norm-matched ordinary cooperation: a 26.4% reduction, 11/12 wins, and Holm p=0.041992p=0.041992 across the two planned comparisons.

We then restore identical complete states at K=3501K=3501 and run 1500 frozen updates without coordinate proposals. The resulting IGDs are:

Frozen updateReturned-set IGD ↓
Independent STCH with common archive/readout0.041243
Norm-matched ordinary cooperation0.035388
Orthogonal cooperation0.029413

The 16.9% gain over ordinary cooperation remains after formation history is shared. Moreover, F7's ρ=0.5\rho=0.5 cap already gives the ordinary input a strict first-order descent margin. Orientation therefore contributes beyond preventing an otherwise non-descending input. Section 4.1 and Appendix F.8, Table 17, now present these interventions and distinguish their returned-set effect from current-particle quality. Norm matching uses the orthogonal candidate norm at each branch's current state.

W2 & Q2: Does Adam preserve the claimed descent property in its actual parameter update, and does that property improve learning after freezing?

Response: We directly measured a 2 × 3 Adult experiment: Adam or direct Armijo, each with orthogonal, norm-matched ordinary, or removed cooperation, from the same 46-epoch state. Six fresh seeds run 30 frozen epochs. All 36 primary validation selections occur 1–10 epochs after freezing.

For orthogonal inputs, Adam gives positive g⊤Δxg^\top\Delta x in 8.42% of updates and same-batch scalar increases in 11.15%. Armijo gives 0% for both. Thus, input orthogonality and actual Adam descent differ. The selected-test orthogonal-minus-ordinary HV gains are +0.001043 and +0.001077 under Adam and Armijo, both with Holm p=0.65625p=0.65625. Fixed-endpoint gains are +0.002022 and −0.000724. This control identifies where the geometric property enters the implemented update: direct Armijo realizes batch descent, and the paired held-out effects are small. Section 4.2 and Appendix F.14, Table 24, now report both readouts and the actual displacements.

W3 & Q3: The benchmark selection, decision dimensions, and objective counts leave the method's broader applicability unclear.

Response: The original problems were selected to represent continuous and disconnected fronts and nonlinear decision-space geometry. We now test cooperative direction across all seven original tasks and d=10,30,100d=10,30,100 for ZDT3/F7. F9 gives IGD 0.009107 versus 0.066539, with 12/12 wins and Holm p=0.005371p=0.005371 across the 11-setting family. We also add DTLZ2 at M=2,3,5M=2,3,5, fixing d=30d=30, 200 particles, and 1001 updates. Appendices F.9 and F.13, Tables 18 and 23, report the complete dimension and objective-count results. Direction effects vary across these fixed configurations; the F9 result provides the clearest additional evidence, while DTLZ2 extends the implementation to five objectives.

W4 & Q3: Can the learning evidence extend beyond small tabular networks to different tasks and larger models?

Response: We now include dense-prediction experiments on NYUv2 and Cityscapes in Section 4.2 and Appendix F.16, Table 27. NYUv2 jointly predicts semantic segmentation, depth, and surface normals; Cityscapes jointly predicts segmentation and inverse depth. All three methods use six models with the same ResNet18 encoder and shared 128-channel decoder, so these comparisons extend the evaluation to pixel-level multitask prediction.

MethodNYUv2 test HV ↑Cityscapes test HV ↑
MOGF/D0.197910 ± 0.0006120.545621 ± 0.001042
ParetoMTL0.180049 ± 0.0081400.536945 ± 0.000860
Particle-WFR0.182915 ± 0.0015740.541501 ± 0.001006

Thank you for the additional experiments. F7 and F9 strengthen the evidence for orthogonal cooperation, and the new learning experiments broaden the evaluation.

The benchmark suites are designed to cover different problem characteristics, so the choice of instances matters. The new synthetic experiments mainly vary the decision dimension or the number of objectives, and it remains unclear why these particular subsets were chosen and other relevant problems were omitted. This makes it difficult to judge how representative the reported gains are.

On Adult, the results vary with checkpoint selection, leaving open whether preserving descent improves test performance.

Given these limits on the demonstrated scope of the mechanism's contribution, I'm keeping my rating at 2.

AuthorsOpenReview ↗

Response by the authors (1/4)

Thank you for your follow-up comment and for recognizing that F7, F9, and the added learning experiments strengthen the evaluation. Below please find our responses, with core experimental tables embedded directly in each comment.

Point 1 (Part A): Complete Suite Coverage (All 21 Continuous Problems) & Mechanism Analysis

The benchmark suites are designed to cover different problem characteristics, so the choice of instances matters. The new synthetic experiments mainly vary the decision dimension or the number of objectives, and it remains unclear why these particular subsets were chosen and other relevant problems were omitted. This makes it difficult to judge how representative the reported gains are.

Thank you for pointing this out. To eliminate subset-selection ambiguity, we have now evaluated all 21 continuous problems across ZDT1–6, DTLZ1–7, and LZ09-F1–9 (ZDT5 uses binary bit-strings outside continuous gradient flow). We ran 17 canonical problems over 12 paired seeds each (ZDT4, ZDT6, DTLZ1–6, and all nine LZ09-F1–F9 under official jMetal grouping), reporting the original six-seed ZDT1–3 and DTLZ7 panel separately in Comment (2/4). Every new run transfers fixed hyperparameters from its existing family (ZDT1, ZDT2, DTLZ7, or non-coordinate F9; N=50,T=5001N{=}50,T{=}5001 for M=2M{=}2; N=200,T=1001N{=}200,T{=}1001 for M=3M{=}3), with no per-problem retuning and no F7 coordinate proposals.

Table 1. Returned-set IGD ↓\downarrow (mean±SD, 12 paired seeds) varying only the projected cooperation channel. †: Orthogonal beats both controls under exact two-sided sign-flip tests (40964096 signs) with 34-contrast Holm p<0.05p{<}0.05 (*: Removed beats Orthogonal at pHolm<0.05p_{\text{Holm}}{<}0.05).

Task (d,Md,M)OrthogonalMatchedRemovedHolm pp
ZDT4 (10,2)26.381±8.85325.162±9.14626.751±8.7301.00/1.00
ZDT6 (10,2)0.0222±0.01770.0189±0.01940.0192±0.01931.00/1.00
DTLZ1 (30,3)40.802±4.15041.053±4.86742.224±4.4541.00/1.00
DTLZ2 (30,3)0.0470±0.02560.0454±0.02610.0454±0.02600.422/0.432
DTLZ3 (30,3)131.88±14.95130.53±15.20127.40±14.801.00/0.938
DTLZ4 (30,3) †0.3900±0.06680.6540±0.09880.6556±0.09900.017/0.017
DTLZ5 (30,3)0.0278±0.01170.0278±0.01170.0278±0.01171.00/1.00
DTLZ6 (30,3)17.818±0.84917.728±0.76617.784±0.8221.00/1.00
LZ09-F1 (30,2) *0.00787±0.000050.00792±0.000130.00773±0.000061.00/0.017*
LZ09-F2 (30,2) †0.00993±0.001500.05724±0.008310.03278±0.016840.017/0.017
LZ09-F3 (30,2) †0.00872±0.001830.02344±0.004580.01612±0.004950.017/0.020
LZ09-F4 (30,2) †0.00790±0.000080.01137±0.000990.00977±0.003180.017/0.017
LZ09-F5 (30,2) †0.00924±0.002100.02325±0.004700.01531±0.004940.017/0.017
LZ09-F6 (10,3)0.08395±0.011620.08205±0.015960.09184±0.012901.00/1.00
LZ09-F7 (10,2) †0.8100±0.26291.3352±0.40101.6025±0.34450.046/0.017
LZ09-F8 (10,2) *0.4406±0.07950.3598±0.11400.1678±0.04790.594/0.017*
LZ09-F9 (30,2) †0.00902±0.001670.06817±0.010740.02969±0.004320.017/0.017

Why orthogonal projection helps on specific problem structures:

  • Nonlinear PS linkage & density bias (DTLZ4, LZ09-F2–F5, F7, F9 — 7 dual-control wins): When the Pareto set curves nonlinearly across coordinates (LZ09 trigonometric/power linkages) or compresses density (DTLZ4, xi100x_i^{100}), inter-particle coverage forces acquire strong projections along the scalarized gradient gg. Unprojected (Matched) cooperation repeatedly perturbs normal convergence and pulls particles off their preference rays, whereas Orthogonal projection decouples tangential spreading from normal descent.
  • Separable or Griewank landscapes (neutral & negative boundaries): On axis-separable/degenerate problems (ZDT4/6, DTLZ1–3/5/6), position and distance variables already lie in orthogonal coordinate blocks, making projection neutral. Conversely, on smooth power-law F1 and Griewank F8 (dense local basins where frozen lateral pushing dislodges particles across narrow ridges), disabling frozen cooperation is preferable. Thank you very much.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

Dear Reviewers and Area Chair,

Thank you for your reviews; below are our Nov 14 revision, supplementary code archive, and second-round analyses for the final manuscript.

Theoretical guarantees and related-method comparison (sSk6, qTsf, CNtg). We prove finite-step backtracking for Eq. (16) (Proposition D.2), analyze formation-stage drift versus frozen-preference stationarity (Table 19; second-round Section 3 derivation), and contrast exact orthogonality between transport and descent directions against MOO-SVGD, PNG, and MGDA/D (Appendix A).

Isolating orthogonal cooperation and full benchmark expansion (sSk6, DotQ, CNtg). Matched-norm and shared-population ablations isolate orthogonal projection from unprojected and scalarization baselines alongside high-dimensional and many-objective scaling (Appendices F.8–F.13) and zero-retuning evaluation across the full ZDT, DTLZ, and LZ09 suites.

Multi-task learning benchmarks and optimizer controls (DotQ, CNtg). We add NYUv2, Cityscapes, and a larger MultiMNIST CNN (Appendices F.15–F.16), plus Adult and MSLR-WEB10K controls separating step-size monotonicity from checkpoint selection.

Other revisions (qTsf, DotQ, CNtg). Appendix B.1 adds Algorithms 1–3, Appendix F.7 reports sensitivity and Optuna logs, and we clarify vector-zero notation and method scope.

Thank you for the constructive discussion.

Best regards,

Authors of Submission 41554

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.