Bayesian optimization (BO) over discrete spaces—such as molecular structures or graph configurations—is notoriously difficult because these domains lack gradient information and valid candidates are sparse. A new preprint on arXiv proposes a strategy that combines latent-space BO with acquisition-function-guided generation to overcome these obstacles.

The method fits a surrogate model in the latent space of a generative model, bypassing the need for gradients and the problem of sparse valid solutions. Instead of generating candidates randomly, the acquisition function steers the generative model toward regions of the latent space that are likely to yield high-performing discrete objects.

This summary is based solely on the preprint's abstract (arXiv:2610.22763). Since only one source is available, there are no independent results to compare or contrast, and the findings have not yet been peer-reviewed.