Fixed-Dimensional Latent Flow Model Generates Variable-Size 3D Molecules

Researchers present a transformer-based autoencoder that maps 3D molecules of differing sizes into a single fixed-dimensional latent space, removing the need to fix molecule size in advance. This matters because molecular size is tied to composition, structure and other target properties in drug and materials discovery. The approach is described as equivariant-free, meaning it avoids the geometric constraints commonly used in 3D molecular generators.