The Multi Modal Models Working Group is delighted to welcome Henry Moss of Lancaster University to discuss the impact Gen AI could have on how we carry out scientific experiments.

Abstract: Generative AI is poised to reshape how scientific experiments are conceived and iterated. Yet despite the ability of popular models like diffusion and flow matching to sample from complicated data distributions, one key bottleneck remains: effective experimental design requires more precise control than “unconditional" sampling. The central challenge is finding outputs that are both likely under the generative model but also meet experiment-specific constraints.

We present a framework for extracting low-dimensional, Euclidean latent spaces from any generative model without additional training of the model. The spaces enable standard experimental design and constrained optimisation methods to directly traverse the model’s output manifold. Our approach is architecture-agnostic, computationally lightweight, and generalises across modalities — from images and audio to video and structured objects like proteins.