NEAT: A New Autoregressive Transformer for 3D Molecular Generation
A new model named NEAT (Neighborhood-Guided, Efficient, Autoregressive Set Transformer) has been created by researchers to generate 3D molecular structures. Unlike earlier autoregressive models that depend on fixed atom orderings, which disrupt permutation invariance, NEAT considers molecular graphs as collections of atoms. It establishes an order-agnostic distribution for permissible tokens at the graph's boundary, thus maintaining atom-level permutation invariance. This model demonstrates cutting-edge generation quality on the QM9 and GEOM-Drugs datasets, providing an effective alternative to methods based on diffusion and flow matching. The findings are presented in arXiv:2512.05844v3.
Key facts
- NEAT stands for Neighborhood-Guided, Efficient, Autoregressive Set Transformer.
- It is designed for 3D molecular generation.
- The model ensures permutation invariance by treating molecular graphs as sets of atoms.
- It learns an order-agnostic distribution over admissible tokens at the graph boundary.
- NEAT achieves state-of-the-art generation quality on QM9 and GEOM-Drugs datasets.
- It offers an efficient alternative to diffusion- and flow-matching-based approaches.
- The research was published on arXiv with ID 2512.05844v3.
- The approach overcomes limitations of canonical atom orderings used in prior works.
Entities
Institutions
- arXiv