Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation
Cai Zhou, Zijie Chen, Zian Li, Jike Wang, Kaiyi Jiang, Pan Li, Rose Yu, Muhan Zhang, Stephen Bates, Tommi Jaakkola · 2026
Read Rethinking Diffusion Models with Symmetries through Canonicalization with Applications to Molecular Graph Generation by Cai Zhou, Zijie Chen, Zian Li, Jike Wang, Kaiyi Jiang, Pan Li, Rose Yu, Muhan Zhang, Stephen Bates, Tommi Jaakkola (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15022.txt.
Description
Zhou and colleagues rethink diffusion models with symmetries through canonicalization. Many patterns in chemistry stay the same when parts are swapped or turned. The usual response bakes that sameness into the design with equivariant denoisers and invariant priors. This paper instead maps each sample to one standard form, trains a plain model on that standard slice, and restores the sameness by sampling a random symmetry at generation time.
The team proves the method sound and able to match any invariant target, with more power than rivals. This speeds training by cutting score complexity from group mixtures and variance in flow matching. Aligned priors and optimal transport add further gains. For 3D molecule making, the method beats prior rivals, and the CanonFlow design leads on the GEOM-DRUG set, even with few steps.
Catalog record
| CREATOR | Cai Zhou, Zijie Chen, Zian Li, Jike Wang, Kaiyi Jiang, Pan Li, Rose Yu, Muhan Zhang, Stephen Bates, Tommi Jaakkola |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 112 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-15022