Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization
About
Offline black-box optimization aims to discover novel designs with high property scores using only a static dataset, a task fundamentally challenged by the out-of-distribution (OOD) extrapolation problem. Existing approaches typically bifurcate into inverse methods, which struggle with the ill-posed nature of mapping scores to designs, and forward methods, which often lack the distributional expressivity to quantify uncertainty effectively. In this work, we propose SPADE (Support-Proximity Augmented Diffusion Estimation), a novel framework that reimagines forward surrogate modeling through the lens of conditional generative modeling. SPADE models the forward likelihood p(y|x) using a diffusion model, but with two critical enhancements to tailor it for optimization: (1) a Calibrated Diffusion Estimation module that enforces global consistency in statistical moments and pairwise rankings, and (2) a Support-Proximity Regularization mechanism that implicitly internalizes the data manifold constraint p(x) via kNN-based density estimation. Theoretically, we prove that our regularization is first-order equivalent to maximizing a Bayesian posterior with a valid design prior. Empirically, SPADE achieves state-of-the-art performance across Design-Bench tasks and an LLM data mixture optimization benchmark.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Offline Black-box Optimization | Ant | Normalized Median Score0.935 | 25 | |
| Offline Black-box Optimization | TF8 | Normalized Median Score67.9 | 25 | |
| Offline Black-box Optimization | TF10 | Normalized Median Score0.748 | 25 | |
| Offline Black-box Optimization | SuperC | Normalized Median Score43.6 | 25 | |
| Offline Black-box Optimization | LLM-DM | Normalized Median Score99.4 | 25 | |
| Offline Black-box Optimization | D'Kitty | Normalized Median Score0.905 | 25 | |
| Offline Black-box Optimization | Overall Task Suite SuperC, Ant, D’Kitty, LLM-DM, TF8, TF10 | Mean Rank1.7 | 24 |