Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization

ICML 2026 (Seoul, South Korea) · also accepted to the ICLR 2026 DeLTa workshop · my role: co-first author.

Paper (arXiv) Project page Code Slides (PDF) Slides (PPTX) Poster

TL;DR

Offline black-box optimization searches for high-scoring designs from a fixed dataset with no online oracle. The central failure mode is out-of-distribution exploitation — optimizers chase surrogate errors in regions the data never covered. SPADE turns forward surrogate modeling into a calibrated conditional diffusion problem and injects a kNN support-proximity prior that shrinks predicted means and inflates uncertainty in low-density regions, keeping the search both expressive and conservative.

Contributions

• Conditional diffusion surrogate that models the forward likelihood \(p_\theta(y\mid x)\), yielding a predictive distribution rather than a point estimate.
• Calibrated diffusion estimation via moment matching and pairwise rank consistency, so the surrogate is actually useful for acquisition optimization.
• Support-proximity regularization using kNN density — which we prove is equivalent to Bayesian inference under a valid design prior.
• State-of-the-art results on Design-Bench and language-model optimization: mean rank 2.8/24 and top-2 finishes on 5 of 6 tasks by normalized max score.

Slides

Can't see the slides? Download the PDF or the PPTX.

Poster

Download the poster (PDF).

BibTeX

@inproceedings{yang2026spade,
  title     = {Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization},
  author    = {Yang, Yonghan and Yuan, Ye and Sun, Zipeng and Du, Linfeng and
               He, Bowei and Wu, Haolun and Chen, Can and Liu, Xue},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {306},
  address   = {Seoul, South Korea},
  publisher = {PMLR},
  year      = {2026}
}

Download .bib (includes the arXiv entry).

Work supervised by Ye Yuan and Prof. Xue (Steve) Liu at Mila – Quebec AI Institute.

Recommended citation: Yonghan Yang*, Ye Yuan*, Zipeng Sun, Linfeng Du, Bowei He, Haolun Wu, Can Chen, and Xue Liu. (2026). "Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization." Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306, Seoul, South Korea. (* equal contribution)
Download Paper | Download Slides | Download Bibtex