Technical Summary: ParetoPilot

Problem Statement

Offline Multi-Objective Optimization (Offline MOO) seeks to discover novel Pareto-optimal designs using only static, pre-collected datasets, avoiding expensive online environment interactions. This is critical in domains like drug discovery, material design, and hardware architecture search where function evaluations are costly or restricted.

The core challenge lies in simultaneously achieving convergence (pushing solutions toward the unobserved true Pareto Front) and diversity (ensuring broad coverage across the objective trade-off space). Current state-of-the-art approaches predominantly rely on external surrogate models (e.g., proxy regressors or preference classifiers) to guide generative processes. This reliance introduces three significant bottlenecks:

1. Computational Overhead: Training separate surrogates for each objective requires extensive hyperparameter tuning and resources.
2. Unreliable Extrapolation: External proxies often produce deceptive scores for out-of-distribution (OOD) designs, leading to manifold collapse and invalid solutions.
3. Privacy and Deployment Constraints: Training task-specific proxies requires continuous access to the original offline dataset, violating the foundation model paradigm of deploying models solely as weights and conflicting with strict data privacy regulations.

Methodology: ParetoPilot and the IPG Engine

The authors propose ParetoPilot, a zero-surrogate diffusion framework that eliminates the need for external auxiliary models. Instead, it leverages the conditional priors inherently embedded within a pre-trained conditional diffusion model. The framework operates by interleaving a novel Infer-Perturb-Guide (IPG) engine within the unconditional denoising steps of the reverse generation process.

The IPG engine executes three sequential operations at specific intervals during the diffusion process:

1. Infer: Implicit Manifold Anchoring

Rather than using an external regressor, the framework implicitly infers the instantaneous objective direction ($y^*_t$) for a noisy sample $x_t$.

• Mechanism: It minimizes an alignment loss ($L_{align}$) that measures the distance between the conditional noise prediction ($\epsilon_\theta(x_t, y, t)$) and the unconditional noise prediction ($\epsilon_\theta(x_t, \emptyset, t)$).
• Stabilization: To counteract the high-frequency stochasticity of noise predictions at large timesteps, the method employs a global, un-reset Adam optimizer. By maintaining the optimizer's momentum state across diffusion steps, it acts as a natural low-pass filter, smoothing the optimization trajectory of the inferred anchor $y^*_t$.

2. Perturb: Orthogonal Fluid Dynamics

Once the anchor $y^*_t$ is established, the engine synthesizes a perturbation vector to guide the batch of candidates toward the Pareto Front while maintaining diversity. This is achieved through a mathematically orthogonalized fluid dynamics system:

• Warmup Phase: Perturbations are confined to steps where $t \le T - T_{warmup}$ to allow the anchor to stabilize and prevent early manifold collapse.
• Convergence Force (Parallel Gravity): A unified parallel gravity field ($\vec{d}_{conv}$) drives all solutions toward optimal regions while preserving their relative lateral distances, thus retaining the prior diversity of the unconditional distribution.
• Diversity Force (Edgeness-Aware Repulsion): To ensure coverage of extreme trade-off designs, an "edgeness" multiplier ($E_i$) is applied to solutions near boundaries, amplifying a repulsive force ($\vec{F}_{rep}$) that pushes them outward.
• Orthogonalization: To prevent repulsive forces from opposing convergence, the repulsive vector is projected onto the orthogonal complement of the gravity field using Gram-Schmidt orthogonalization. This strictly decouples exploration (diversity) from exploitation (convergence).
• Dynamic Annealing: A temporal schedule blends the convergence and diversity forces, transitioning from convergence-heavy to diversity-heavy (or vice versa) as the process approaches $t=0$.

3. Guide: CFG Extrapolation

The perturbed target condition ($y^{target}_t$) is fed directly into the standard Classifier-Free Guidance (CFG) formulation.

• The modified noise prediction $\tilde{\epsilon}_t$ is computed as:
  $$\tilde{\epsilon}_t = \epsilon_\theta(x_t, \emptyset, t) + w_t \cdot (\epsilon_\theta(x_t, y^{target}_t, t) - \epsilon_\theta(x_t, \emptyset, t))$$
• The guidance scale $w_t$ is linearly interpolated alongside the perturbation schedule. This allows the model to steer the generation toward diverse, Pareto-optimal regions without ever leaving the valid data manifold defined by the unconditional distribution.

Key Contributions

1. Zero-Surrogate Framework: ParetoPilot is the first framework to perform offline MOO by directly leveraging intrinsic conditional information from pre-trained diffusion models, eliminating the need to train auxiliary proxy models.
2. IPG Engine: A novel mechanism that dynamically alters generative trajectories by inferring implicit conditions and applying orthogonalized fluid forces. This explicitly guides the denoising process to approximate the true Pareto front while preserving data privacy.
3. Plug-and-Play Capability: The method acts as a module that can be deployed on learned distributions without access to original training data, aligning with practical downstream deployment scenarios.

Experimental Results

The authors evaluated ParetoPilot on 51 continuous offline MOO tasks across four domains: Synthetic Functions, Multi-Objective Neural Architecture Search (MO-NAS), Scientific Design (SciDesign), and Real-world Engineering (RE).

• Performance: ParetoPilot achieved an Average Rank of 6.81 and a 10.46% improvement in Hypervolume (HV) across all tasks, outperforming 14 state-of-the-art baselines (including surrogate-based forward methods like IOM, PcGrad, and generative inverse methods like ParetoFlow and PGD-MOO).
• Robustness: Unlike surrogate-based methods that suffered severe performance degradation (and negative HV improvements) on complex discrete/structural tasks (MO-NAS, SciDesign) due to OOD evaluation errors, ParetoPilot maintained robust performance. The unconditional diffusion steps inherently restricted trajectories to valid data manifolds, preventing the synthesis of nonsensical molecules or broken architectures.
• Ablation Studies: Removing key components confirmed their necessity:
  • Without Smoothing: Performance dropped significantly (Rank 3.08) due to trajectory oscillations caused by unfiltered noise.
  • Without Edgeness: The system failed to efficiently discover extreme trade-off designs.
  • Without Orthogonalization: Convergence and diversity forces mutually neutralized, reducing overall HV improvement.

Significance and Claims

The paper claims that ParetoPilot represents a paradigm shift in Offline MOO by resolving the tension between data privacy, computational efficiency, and optimization reliability.

• Privacy Preservation: By removing the need for external surrogates, the method strictly preserves data privacy, as it does not require re-accessing the original offline dataset for proxy training.
• Reliability: It avoids the "deceptive evaluations" common in surrogate-based methods, which often lead to manifold collapse when extrapolating to OOD regions.
• Deployment Viability: The approach aligns with the foundation model paradigm, allowing models to be shared solely as weights while still enabling complex optimization tasks downstream.

The authors acknowledge limitations, noting that the generation quality is inherently dependent on the base diffusion model's ability to capture the initial data distribution and that the IPG engine introduces additional inference latency compared to simple forward surrogate evaluations. However, they posit that the trade-off is justified by the elimination of proxy training overhead and the robustness of the results.