SPADE — Diffusion Surrogates for Offline Optimization

Support-Proximity Augmented Diffusion Estimation (SPADE) models the forward likelihood \(p(y\mid x)\) with a calibrated conditional diffusion model, then adds a k-nearest-neighbour support prior so an offline optimizer cannot exploit regions the data never covered. It reaches state-of-the-art results on Design-Bench and language-model optimization.

• Venue: ICML 2026 (co-first author) · ICLR 2026 DeLTa workshop
• Paper: arXiv:2605.11246
• Code: github.com/HarryYoung2018/spade
• Project page: harryyoung2018.github.io/spade
• Stack: Python, PyTorch, diffusion models, evolutionary search