Title: ParetoPilot: Achieving Zero-Surrogate Offline Multi-Objective Optimization Through Infer-Perturb-Guide Diffusion

Abstract:

Offline multi-objective optimization (Offline MOO) seeks to identify novel Pareto-optimal designs using static datasets, thereby avoiding the high costs associated with extensive environment interactions. Although recent generative approaches have shown promise, they typically depend on external surrogate models. This reliance creates substantial computational burdens, leads to misleading evaluations, and diverges from the standard practice of jointly training primary generative models with conditional inputs. To overcome these limitations, we introduce ParetoPilot, a novel diffusion-based framework for offline MOO that operates without surrogates. ParetoPilot capitalizes on the conditional priors already embedded in pre-trained diffusion models. Central to this framework is the Infer-Perturb-Guide (IPG) engine, which is integrated directly into the unconditional denoising steps of the reverse generation process. Initially, the IPG engine implicitly determines the instantaneous objective direction by comparing conditional and unconditional noise predictions. Subsequently, it constructs a dynamically annealed perturbation vector by mathematically orthogonalizing a parallel gravity field—ensuring strict convergence—and an edgeness-aware repulsive force that promotes mutual diversity. This perturbed target then guides the generation process through standard Classifier-Free Guidance (CFG). Our extensive experiments across 51 tasks reveal that ParetoPilot surpasses 14 leading surrogate-based and inverse generative baselines. By removing the need for auxiliary proxy training, the method not only safeguards data privacy but also delivers superior hypervolume improvements and robust coverage of the Pareto front.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC