Research map / Image generation / Autoregressive and latent generation

NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Equivariant diffusion for 3D molecules · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NEAT, an order-agnostic autoregressive set transformer for unconditional 3D molecule generation trained with neighborhood-guided subgraph splitting and a flow-matching coordinate head, matching flow-matching models on GEOM while running several times faster.

Problem

Autoregressive transformers for 3D molecular generation require canonical atom orderings, which are not inherently defined for molecular graphs and bias next-token prediction toward ordering conventions. The paper targets permutation-invariant autoregressive 3D generation.

Approach

Train a bidirectional set transformer on source/target subgraph splits sampled from molecular graphs, so that the target is the set of boundary (1-hop neighbor) atoms; atom types are predicted as a mean one-hot distribution over the target set and coordinates via a conditional flow matching head, removing the need for canonical atom ordering.</ph>

Key results

• On QM9, NEAT reaches 95.2% xyz2mol validity and 91.6% xyz2mol uniqueness, outperforming other autoregressive baselines (QUETZAL 93.7% / 89.1%).
• On GEOM, NEAT achieves 99.7% lookup validity, the highest among compared models, with 93.9% xyz2mol validity.
• NEAT is reported to run 5x faster than SemlaFlow and QUETZAL, 16x faster than FlowMol3, and 34x faster than Megalodon-FM.

Surprising findings

• Greedy atom-type sampling maximizes validity but sharply reduces uniqueness, showing a validity-uniqueness trade-off.
• On prefix completion, QUETZAL reaches only 28.1% xyz2mol validity while NEAT reaches 92.7%, because canonical ordering prevents completing arbitrary prefixes.

Limitations

• Neighborhood guidance relies on bond connectivity from the molecular graph, limiting applicability to inorganic crystals or metal-organic compounds.
• Standard transformer is not SE(3)-equivariant; relies on random rotation augmentation.
• Does not extrapolate beyond the maximum atom count seen during training; error propagation possible for larger molecules.

Experimental setup

Datasets: QM9 GEOM-Drugs

Benchmarks: QM9 GEOM-Drugs

Metrics: atom stability molecular stability lookup validity lookup uniqueness xyz2mol validity xyz2mol uniqueness xyz2mol novelty runtime

Baselines: G-SchNet G-SphereNet Symphony QUETZAL Geo2Seq FlowMol3 SemlaFlow Megalodon-FM

Builds on: QUETZAL NanoGPT GPT-2 conditional flow matching SimpleFold

Closest papers by meaning

• Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.88)
• All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.88)
• Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space (ICLR 2025 · similarity 0.88)
• Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.87)
• 3D Molecule Generation from Rigid Motifs via $\mathrm{SE}(3)$ Flows (NeurIPS 2026 · similarity 0.87)
• Learning-Order Autoregressive Models with Application to Molecular Graph Generation (ICML 2025 · similarity 0.87)