Constructing a finite Pareto representation requires allocating a limited number of solutions across the Pareto front (PF). A desired preference distribution specifies this allocation by assigning more solutions to some PF regions and fewer to others. Scalarization generates Pareto solutions by varying objective weights and solving single-objective problems. However, a distribution of scalarization weights generally does not translate into the same distribution of Pareto solutions along the PF, because the weight-to-solution map is nonlinear and unknown in advance.To account for this distortion, we introduce SURFACE (Scalarization Updates for Redistributing Finite Approximations with Chart Estimation}), which adaptively updates scalarization weights to match a desired PF distribution. SURFACE estimates how each region of the weight grid corresponds to an area of the PF from the current scalarized solutions, then redistributes the weights accordingly. The method supports parallel, warm-started, inexact optimization with standard single-objective solvers. Theoretically, we establish that, under the stated regularity conditions, SURFACE achieves linear convergence of the preference distribution mismatch to a finite-resolution error floor of O(N−2/(M−1)), where N is the number of adaptive weights and M is the number of objectives. Empirically, experiments across multiple benchmarks demonstrate improved target-mass balance and PF coverage compared with established baselines.
Figures & tables
Figure 1: Preference ϑ(fPF(w)) and the factor J(w) measuring PF area scaling jointly determine the target weight density ρϑ(w) .
Figure 2: An overview of SURFACE. Cooler regions indicate lower target density, while warmer regions indicate higher density and a denser allocation of Pareto solutions. SURFACE estimates the weight-to-PF geometry and updates the scalarization weights to match this target distribution.
Figure 3: Weight-to-PF geometry. A simplex cell τ maps to a PF region fPF(τ) . The factor J(w) generalizes PF traversal speed to local surface-area scaling, with ∫τJ(w)dw giving the area of the corresponding PF region.
Damping αˉ
Contraction ηK
Inner K
Outer T
When to Use
Θ(1)
O(N−3/(M−1))
O(logN)
O(logN)
Linear/superlinear or p>2
Θ(N−1/(M−1))
O(N−1/(M−1))
O(logN)
O(N1/(M−1)logN)
p=2 / moderate K
Θ(N−3/(2(M−1)))
O(1)
O(1)
O(N3/(2(M−1))logN)
p<2 / fixed K
Table 1: Outer–inner iteration trade-offs under αˉ2ηK=O(N−3/(M−1)) . The K and T columns assume linear inner convergence and α=Θ(αˉ) , respectively. The “When to Use” column identifies preferred damping regimes for sublinear inner solvers with ηK=O(K−p) .
Figure 4: Uniform weighting versus SURFACE with ϑ=1,ϑfocus . Warmer colors indicate higher ϑfocus .
Figure 6
Metric
SMS-EMOA
NSGA-III
LMPFE
Subset
UMOD
Uniform
SURFACE
ZDT2, M=2 , N=9
R^ϑ↓
.480±.018
.589±.000
.881±.374
.491±.000
.812±.047
.514
.397
IGD↓
.039±.000
.038±.000
.048±.013
.039±.000
.041±.000
.037
.036
Fill ↓
.126±.003
.089±.000
.156±.068
.122±.000
.101±.003
.087
.073
ΔHV↓
.069±.000
.076±.000
.097±.025
.075±.000
.084±.001
.078
.074
ZDT4, M=2 , N=9
Table 2: Benchmark results for representative baselines. Entries for stochastic baselines are mean ± standard deviation over five seeds. SURFACE and Uniform are deterministic. Complete results for all ten methods are reported in Appendix E.3 .
Figure 8: SURFACE with ϑ=1,ϑfocus . From left to right: DTLZ2 ( M=3 , N=66 ), DTLZ7 ( M=3 , N=153 ), DTLZ7 ( M=2 , N=20 ), and ZDT3 ( M=2 , N=25 ). Warmer PF colors indicate higher ϑfocus on a shared 0 – 1 scale.
Metric
SMS-EMOA
NSGA-III
MOEA/D
MOEA/D-AWA
UMOD
Uniform
SURFACE
Deep Sea Treasure , M=2 , N=14
Rϑ↓
.943±.379
.838±.279
1.296±.201
1.156±.299
1.138±.371
1.069
.249
IGD↓
.346±.038
.335±.030
.416±.056
.387±.067
.383±.071
.077
.026
Fill ↓
.822±.086
.797±.072
.945±.079
.895±.104
.883±.120
.255
.105
HV ↑
1.031±.178
1.093±.142
.812±.131
.909±.187
.911±.208
1.041
1.043
Resource Gathering , M=3 , N=18
Table 3: MO-Gym policy results. Stochastic entries are mean ± sample standard deviation over five seeds. Uniform and SURFACE are deterministic. Bold marks the best mean in each row.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Definition
Problem and PF chart
M
Number of objectives.
. p. 1
x , X
Decision variable and feasible set.
. p. 1
f=(f1,…,fM)
Vector-valued objective.
. p. 1
z
Utopian Tchebycheff reference, with zm<minx∈Xfm(x) .
. p. 6
w
Scalarization weight.
. p. 6
Appendix
Table 4: Recurring notation guide.
Method
Rϑ
IGD
Fill
ΔHV
Coverage
SURFACE
1.18
1.18
1.64
2.55
1.33
Uniform
3.18
2.27
2.45
4.73
2.64
NSGA-III
5.32
4.91
4.45
5.00
4.89
LMPFE
4.91
5.00
5.09
6.18
5.00
SMS-EMOA
5.00
7.00
7.09
3.91
6.36
MOEA/D-AWA
7.27
6.82
6.55
7.27
6.88
Appendix
Table 5: Mean ranks over 11 ZDT/DTLZ task-dimension groups. Coverage averages the first three PF-representation metrics and excludes hypervolume. Lower is better.
Metric
SMS-EMOA
NSGA-III
LMPFE
Subset
MOEA/D-AWA
MOEA/D
DEA-GNG
UMOD
Uniform
SURFACE
ZDT1, M=2 , N=9
Rϑ↓
.431±.015
.608±.000
.563±.060
.769±.001
.434±.076
.897±.239
.516±.122
.483±.065
.505
.316
IGD↓
.037±.000
.042±.000
.039±.003
.040±.000
.039±.000
.049±.005
.041±.002
.037±.000
.037
.035
Fill ↓
.092±.002
.150±.000
.095±.016
.174±.000
.105±.003
.189±.032
.119±.009
.085±.008
.106
.096
ΔHV↓
.048±.000
.055±.000
.053±.004
.049±.000
.052±.000
.060±.004
.055±.003
.049±.001
.048
.045
Coverage ↓
2.67±.38
8.33±.38
5.33±1.69
8.33±.53
4.33±.62
10.00±.73
7.00±1.22
2.67±.96
4.33
2.00
Appendix
Table 6: Complete results for SURFACE and 9 baselines. Stochastic-method entries are mean ± sample standard deviation over five seeds. Uniform and SURFACE are deterministic and have no standard-deviation terms. Each Coverage row reports the mean rank over Rϑ , IGD, and fill distance.
Scalarization is widely used in multi-objective optimization owing to its simplicity and scalability. In many applications, the goal is to generate solutions that represent diverse user preferences, ideally with uniform coverage of the Pareto front (PF). However, uniformly sampling scalarization weights usually induces non-uniform coverage of the PF. We explain this mismatch through the geometry of the scalarization path. As the weight varies, the corresponding solutions traverse the PF at a generally nonuniform speed. Integrating this speed gives cumulative arc length. After normalization by the total length, it defines a cumulative distribution function (CDF) on weights whose inverse selects points at equal arc-length intervals. Building on this insight, we propose SURF (Sampling Uniformly along the Pareto Front). For structured problems, including bi-objective bandits, we derive analytical expressions for this CDF map and the resulting PF-aware weight sampling rule. For general problems, SURF alternates between CDF reconstruction and weight sampling. Theoretically, we show that under provable conditions for bi-objective LS, SURF's CDF reconstruction error converges linearly to an O(N−2) discretization floor. Empirically, experiments on bandits, multi-objective-gymnasium, and multi-objective LLM alignment demonstrate that SURF efficiently achieves more uniform PF coverage than baselines.
Liuyuan Jiang, Chentong Huang, Lisha Chen
Department of Electrical and Computer Engineering University of Rochester, Rochester, NY 14627
Offline multi-objective optimization requires not only moving the objective vectors of candidate designs toward the Pareto front, but also distributing them effectively along it. Generative methods have recently emerged as a natural approach because they learn a distribution over feasible designs while allowing generation to be steered toward promising designs. Existing methods, however, largely retain classical sample-wise guidance strategies, leaving the distribution-level modeling capability of generative methods underused. We propose ParetoTransport, a training-free guidance method for pre-trained flow-matching models that explicitly specifies and refines a population-level distribution in objective space. ParetoTransport guides a flow-matching sampler to iteratively transport the empirical offline distribution toward the Pareto front, with Wasserstein matching to intermediate proxy distributions. This directly controls distributional displacement and mass allocation along the front. We establish a convergence result and demonstrate state-of-the-art performance on standard offline MOO benchmarks, extending recent evaluations beyond hypervolume to generational distance, inverted generational distance, and Wasserstein distance.
Stephanie Holly, Sepp Hochreiter, Werner Zellinger
LIT AI Lab and Institute for Machine Learning JKU Linz, Austria
Multi-objective optimization involves diverse preferences over conflicting objectives. Rather than returning a finite set of trade-off solutions, preference-conditioned Pareto set learning seeks to map a specified preference directly to a corresponding Pareto optimal solution. Yet in many real-world scenarios, optimization must proceed offline without further objective evaluations, leaving only a limited static dataset that sparsely covers the decision and objective spaces. Existing methods either optimize separately for each preference or train a problem-specific Pareto solution model, requiring repeated search across preferences or repeated training across problems. We introduce in-context preference-conditioned Pareto set learning (ICPSL), a framework that amortizes both preference-wise solution search and problem-specific model construction. ICPSL pretrains a foundation-style hypernetwork on a large and diverse set of synthetic prior data. Given the offline observations of an unseen problem, the pretrained hypernetwork learns to generate a problem-specific, preference-conditioned Pareto solution model in a single forward pass. The generated model can then be reused to predict solutions for arbitrary preferences without problem-specific training or preference-wise search. Experiments on synthetic and real-world problems show that ICPSL produces high-quality Pareto solutions while providing substantial speedups in problem-specific solution-model construction.