Generative Bayesian Optimization (GenBO) at ICLR 2026
We published GenBO at ICLR 2026:
Generative Bayesian Optimization: Generative Models as Acquisition Functions
Bayesian optimization is a technique for finding the best setting for something (like a chemical formula or an engineering design) using as few costly trial-and-error experiments as possible. This paper shows how to use generative AI models, similar in spirit to the technique behind tools like ChatGPT’s preference tuning, to directly generate promising candidates to try next, rather than building a separate prediction model first. This makes it possible to search much larger batches of candidates at once, including in tricky spaces with many dimensions or non-numeric options.