Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (DiBO)

DiBO introduces a novel framework that combines diffusion models with ensemble-based uncertainty quantification to solve high-dimensional black-box optimization problems. By iteratively training diffusion models to capture data distributions and fine-tuning them for posterior inference, our method effectively balances exploration and exploitation in high-dimensional spaces, outperforming existing approaches across various synthetic and real-world tasks.

For a visual explanation of the DiBO framework, please see the image below:

You can find the paper on OpenReview and the code on GitHub.

Recommended citation: Yun, T.*, Om, K.*, Lee, J., Yun, S., & Park, J. (2025). "Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization." ICML; FPI @ ICLR Workshop.
Download Paper