NEAT reorders autoregressive 3D molecule generation around atom neighborhoods
The authors of the NEAT paper say their neighborhood-guided training removes the need for canonical atom orderings while matching or beating diffusion and flow baselines on quality and speed.
In plain words
A team proposes NEAT, a model that generates 3D molecular structures one atom at a time. Standard sequence models need a fixed order of tokens, but molecules have no natural ordering, so earlier methods imposed arbitrary conventions. NEAT instead treats the molecule as an unordered set of atoms and predicts from each atom's neighborhood.
What was announced
Autoregressive transformers generate tokens in sequence, which suits text but not molecules, where atoms form a graph with no inherent ordering. The NEAT authors note that prior work relied on canonical atom orderings, which they say breaks permutation invariance and biases next-token prediction toward those conventions. Their neighborhood-guided training strategy instead treats molecular graphs as sets of atoms and learns an order-agnostic distribution over admissible tokens at the graph boundary, which they report yields atom-level permutation invariance.
According to the abstract, NEAT achieves state-of-the-art generation quality on the QM9 and GEOM-Drugs benchmarks, can complete partial molecular structures from a given prefix, and runs significantly faster than existing baselines, positioning autoregressive models as an efficient alternative to diffusion- and flow-based 3D molecule generators. These are the authors' own reported results; the abstract does not include the underlying measurements, and independent evaluation is not part of this evidence.
What it does not show
- Evidence is limited to the paper's abstract; benchmark numbers, baselines and speed figures are asserted but not shown, and no independent confirmation is available.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.