Riemannian MeanFlow
Dongyeop Woo, Marta Skreta, Seonghyun Park, Kirill Neklyudov, Sungsoo Ahn · 2026
Read Riemannian MeanFlow by Dongyeop Woo, Marta Skreta, Seonghyun Park, Kirill Neklyudov, Sungsoo Ahn (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-2602-07744.txt.
Description
Generative models on curved spaces, used for protein shapes and DNA design, usually need dozens of neural-network evaluations to produce one sample. Dongyeop Woo and four co-authors introduce Riemannian MeanFlow, a method that learns flow maps directly on manifolds so good samples can emerge from as little as one forward pass.
The paper derives three equivalent descriptions of the manifold's average velocity and studies how to parameterize and stabilize training on high-dimensional curved spaces. In promoter DNA design and protein backbone generation, the method matches earlier quality with up to ten times fewer evaluations, and few-step maps also enable reward-guided design. It promises faster sampling for scientific workflows that generate thousands of candidates.
Catalog record
| CREATOR | Dongyeop Woo, Marta Skreta, Seonghyun Park, Kirill Neklyudov, Sungsoo Ahn |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 131 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-07744