[Paper] ParetoPilot: Zero-Surrogate Offline Multi-Objective Optimization via Infer-Perturb-Guide Diffusion

Source: arXiv - 2606.04468v1

Overview

Offline multi‑objective optimization (MOO) seeks new designs that simultaneously excel on several metrics—think faster CPUs that also consume less power—using only a static dataset, without costly simulations or real‑world trials. The new ParetoPilot framework shows how to harness pre‑trained diffusion models for this task without training any surrogate evaluator, dramatically cutting compute, preserving data privacy, and delivering stronger Pareto fronts than prior methods.

Key Contributions

• Zero‑surrogate diffusion: Introduces a diffusion‑based optimizer that never builds an auxiliary proxy model, sidestepping the heavy training and potential bias of surrogate‑based pipelines.
• Infer‑Perturb‑Guide (IPG) engine: A three‑step plug‑in that (1) infers the instantaneous objective gradient by contrasting conditional vs. unconditional noise predictions, (2) constructs a mathematically orthogonal perturbation combining a “gravity” pull toward better objectives and an “edgeness‑aware” repulsion for diversity, and (3) feeds this perturbed target into standard classifier‑free guidance (CFG).
• Unified training: Leverages the conditional priors already baked into off‑the‑shelf diffusion models, so no extra condition‑specific fine‑tuning is required.
• Comprehensive evaluation: Benchmarks on 51 offline MOO tasks (covering design, chemistry, robotics, etc.) against 14 state‑of‑the‑art baselines, achieving higher hypervolume and better Pareto front coverage.
• Privacy‑preserving: By eliminating any learned surrogate that could inadvertently memorize the training data, the method respects data confidentiality—a key concern for proprietary design datasets.

Methodology

1. Base diffusion model – Start with a pre‑trained unconditional diffusion model (e.g., a DDPM) that can generate candidate designs from pure noise.
2. Conditional prior extraction – The model already knows how to condition on objectives when trained with classifier‑free guidance (CF‑G). ParetoPilot taps into this latent conditional knowledge without explicit retraining.
3. Infer step – During each reverse diffusion timestep, the model predicts two noise vectors: one unconditional (ε_uncond) and one conditional on a tentative objective direction (ε_cond). Their difference approximates the instantaneous gradient of the multi‑objective loss.
4. Perturb step –
   • Gravity field: A vector pointing toward the region of higher objective values, derived from the inferred gradient.
   • Repulsive field: An “edgeness‑aware” term that pushes the current sample away from already‑generated points, encouraging diversity along the Pareto front.
   • The two fields are orthogonalized (via Gram‑Schmidt) to avoid interference, then combined with a temperature‑like annealing schedule that gradually reduces perturbation strength as generation proceeds.
5. Guide step – The perturbed target replaces the usual CFG target, steering the denoising step toward samples that improve all objectives while staying diverse.
6. Iterate – The process repeats for each diffusion timestep, producing a batch of designs that collectively approximate the Pareto frontier—all without ever training a surrogate evaluator.

Results & Findings

• Across 51 tasks (including circuit layout, molecular property optimization, and robotic gait design), ParetoPilot consistently pushed the Pareto front outward, delivering higher hypervolume scores.
• The orthogonal perturbation proved crucial: ablations removing the repulsive term caused mode collapse (many duplicate designs), while dropping the gravity term slowed convergence.
• Because no surrogate is trained, the total wall‑time dropped by ~40 % compared with the next‑best surrogate‑based method, and the approach scaled linearly with dataset size.

Practical Implications

• Faster design loops – Engineers can plug a pre‑trained diffusion model into their existing offline datasets and start generating Pareto‑optimal candidates instantly, skipping the surrogate‑training stage that can take days.
• Reduced cloud spend – Lower GPU hours translate directly into cost savings for companies running large‑scale design optimization (e.g., semiconductor foundries, drug discovery firms).
• Data confidentiality – Sensitive design data (proprietary CAD files, confidential molecular libraries) never leaves the secure environment, because no external model is trained on it.
• Toolchain integration – The IPG engine is a lightweight add‑on that works with any classifier‑free diffusion backbone (Stable Diffusion, Imagen, etc.), making it easy to embed in existing generative‑AI pipelines or MLOps platforms.
• Multi‑objective “what‑if” exploration – Developers can quickly query different trade‑off surfaces (e.g., latency vs. power vs. area) by adjusting the guidance strength, enabling rapid prototyping and stakeholder negotiation.

Limitations & Future Work

• Dependence on a good diffusion prior – If the base diffusion model was trained on a narrow distribution, ParetoPilot cannot extrapolate far beyond that support.
• Scalability to very high‑dimensional objectives – The current gradient inference scales linearly with the number of objectives; extremely large objective sets may need dimensionality reduction or hierarchical guidance.
• Static datasets only – The method assumes a fixed offline dataset; extending to online or continual learning scenarios (where new data arrives) remains an open challenge.
• Theoretical guarantees – While the orthogonalization ensures non‑interfering forces, formal convergence proofs for the multi‑objective case are still pending.

Future research directions include: (1) integrating adaptive diffusion backbones that can be fine‑tuned on domain‑specific data without breaking the zero‑surrogate premise, (2) exploring hierarchical IPG schemes for dozens of objectives, and (3) coupling ParetoPilot with differentiable simulators to enable hybrid offline‑online optimization loops.

Authors

• Ruiqing Sun
• Sen Yang
• Dawei Feng
• Bo Ding
• Yijie Wang
• Huaimin Wang

Paper Information

• arXiv ID: 2606.04468v1
• Categories: cs.LG, cs.AI, cs.NE, math.OC
• Published: June 3, 2026
• PDF: Download PDF