ADT: All-Atom Diffusion Tokenizer for 3D Molecular Structures

ADT (All-atom Diffusion Tokenizer) is a tokenizer for 3D molecular structures that decouples token granularity from reconstruction resolution, achieving all-atom reconstruction from compact discrete representations. It tokenizes small molecules atom by atom and uses a single token per protein residue that jointly encodes backbone and side-chain geometry, with a diffusion model recovering geometric distortion introduced by tokenization. Experiments show ADT supports high-quality autoregressive generation of small molecules and all-atom protein structures.