acceptodds
Under review as a conference paper at ICLR 2027

Performative Multi-Objective Optimization

Abstract

Performative prediction studies learning under model-induced distribution shift, where the deployed model shapes the very data it aims to predict. Existing theory and algorithms largely focus on a single objective or collapse multiple criteria into one aggregated loss, while practical systems may involve several potentially conflicting objectives such as accuracy, fairness, robustness, and welfare. We introduce performative multi-objective optimization (PMOO), a framework for jointly optimizing several performative objectives, and define Performative Pareto Stationary (PPS) points as solutions satisfying the first-order Pareto stationarity condition under the environment induced by the deployed model. A central challenge in PMOO is the coupling between , induced by the deployed model, and , reflecting evolving trade-offs among objectives. Deployment affects the gradient geometry used to select the multiplier weights, while these weights shape model updates and subsequent environments. For smooth nonconvex objectives, we analyze multi-gradient descent (MGD) and show that its best iterate attains approximate PPS. To stabilize the multiplier dynamics, we propose proximal MGD (PMGD). In the strongly convex setting, we further develop a coupled convergence analysis that makes explicit how performative sensitivity and multiplier geometry influence the dynamics. Together, these results extend performative prediction to a Pareto framework for learning with conflicting objectives.

Primary area:
optimization
Keywords:
Multi-objective optimization, Performative prediction, Multi-gradient descent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ieeWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multiobjective optimization when model deployment changes the data distribution. It defines performative Pareto stationarity (PPS), analyzes repeated multigradient descent (MGD), and introduces proximal MGD to stabilize the combination weights. The theory establishes approximate PPS guarantees for MGD under smooth nonconvex objectives and convergence results under strong convexity and additional geometric conditions. Synthetic experiments and a credit prediction task with simulated strategic responses illustrate the proposed framework.

Critical strengths

The paper addresses an interesting problem: how to optimize multiple potentially conflicting objectives when model deployment changes the data distribution.

Critical weaknesses

The impact of performative feedback on the optimization dynamics could be explained more clearly. Since the proposed methods closely resemble existing gradient methods for multiobjective optimization, the paper would benefit from further discussion of how performative feedback changes their behavior and theoretical guarantees. I would reconsider my score if the authors adequately address this concern.

Questions for authors
  1. Could the authors further justify why PPS is a meaningful target in practical applications and characterize its potential gap from Pareto optimality of the full performative objectives? Since the algorithm freezes the environment when computing gradients, please also highlight which terms in the theoretical bounds capture performative feedback and distinguish the results from standard multi-objective optimization.

  2. Beyond smoother multiplier trajectories, does stabilizing the weights yield measurable improvements in model convergence, robustness to perturbations, or final objective values? A theoretical comparison or targeted experiment would help establish the benefits of PMGD over MGD.

  3. Please clarify the uniqueness of θ⋆\theta^\star, distinguishing uniqueness within a fixed environment from uniqueness of the equilibrium that induces D(θ⋆)D(\theta^\star). In particular, how do the sensitivity assumption on D(θ)D(\theta), multiplier geometry, and joint contraction conditions together establish uniqueness?

Minor corrections:

Line 67: Replace “relecting” with “reflecting.”

Line 117 and elsewhere: Replace “Eq. equation 1” with “Eq. (1)” and correct similar equation references throughout.

Reviewer_Qut3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies performative multi-objective optimization, where deploying a model changes the data distribution underlying multiple objectives. The authors first introduce performative Pareto stationarity, analyze repeated multi-gradient descent (MGD), and propose proximal MGD (PMGD) to stabilize objective weights. They establish approximate stationarity guarantees in the nonconvex setting and convergence under additional strong-convexity and multiplier-geometry assumptions. Experiments use synthetic problems and credit-risk prediction with simulated strategic responses.

Critical strengths
  1. The problem is relevant and interesting. Extending multi-objective optimization to model-dependent environments provides a useful framework for studying conflicting objectives under deployment feedback.

  2. The analysis distinguishes frozen-environment stationarity from stationarity of the full performative objectives.

Critical weaknesses
  1. Limited technical novelty. MGD directly applies standard MGDA, while PMGD adds a temporal proximal penalty. The nonconvex proof appears to combine standard MGDA descent with a distribution-sensitivity bound. Could the authors clarify which technical challenges require substantially new arguments beyond existing MOO and performative prediction analyses?

  2. Limited experimental evidence. The experiments mainly illustrate qualitative behavior. In Figure 1, MGD converges despite abrupt multiplier changes; in Figure 3, the residual curves are similar, and neither method dominates across group losses. A clearer demonstration that multiplier stabilization improves convergence or robustness would strengthen the paper.

  3. Use of main-text space. I suggest fully utilizing the available main-text space to better explain the technical novelty, experimental setup, and interpretation of the results.

Questions for authors

Please check the weakness.

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

This paper extends performative prediction to a multi-objective setting in which the objectives potentially conflict. Starting from a classical MOO method MGDA, the authors introduce an additional term into the optimization objective to mitigate drift in task weights and derive the PMGD. This paper also provide a theoretical analysis of model error. The authors also conduct experiments on synthetic data and Credit data.

Critical strengths
  1. The paper explores the coupling between the data distribution shifts induced by performative prediction and the dynamically changing multipliers in multi-objective optimization. This is a challenging problem.

  2. The paper develops a theoretical framework that connects performative prediction with Pareto stationarity, offering a clear theoretical contribution.

Critical weaknesses
  1. The experiments do not sufficiently demonstrate the practical benefits of the additional regularization term in PMGD. For example, in Figure 4, the curves for MGD and PMGD almost overlap. Similarly, in Figures 1 and 2, MGD achieves nearly the same results as PMGD. Moreover, the improved multiplier stability shown in Figures 1(C1) and 1(C2) does not translate into clearly demonstrated improvements in training time, model parameter stability, or the final PPS residual.

  2. The experimental settings could be more diverse. The evaluation relies mainly on synthetic and semi-synthetic data and lacks experiments involving conflicting objectives in more realistic scenarios, such as accuracy, fairness, robustness, and welfare, as mentioned in the abstract.

Questions for authors
  1. What substantive improvements does PMGD offer over MGD in the experiments? Please provide more direct evidence of these improvements.

  2. To make the empirical evaluation more convincing, please include experiments in a broader range of real-world settings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.