Quotient-Space Diffusion Models for Molecular Generation

A new framework for diffusion-based generative modeling on quotient spaces has been proposed, targeting molecular structure generation with SE(3) symmetry. The approach reduces learning complexity compared to conventional group-equivariant models by eliminating the need to learn the group action component. The sampler guarantees recovery of the target distribution, avoiding heuristic alignment strategies. This work, published on arXiv, formalizes diffusion on general quotient spaces and applies it to 3D molecular generation, a key task in scientific AI.