ReSFlow Enables One-Step Generation for Discrete Data via Statistical Manifolds

1 reports archived1 independent sourcesupdated 11/3/2026, 23:27:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Accepted at NeurIPS 2026, ReSFlow rectifies generation paths on statistical manifolds to enable one-step discrete data generation. By mapping categorical probabilities to a sphere and updating along geodesics, it achieves near-multi-step quality in binary MNIST with a single call and accelerates DNA sequence generation by over 160x.

LATEST/Accepted at NeurIPS 2026, ReSFlow rectifies generation paths on statistical manifolds to enable one-step discrete data generation. By mapping categorical probabilities to a sphere and updating along geodesics, it achieves near-multi-step quality in binary MNIST with a single call and accelerates DNA sequence generation by over 160x.

Hourly heat curve

27 fully observed hours

Now

6.6

Peak

11.811-04

24h change

-43%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

机器之心 (微信公众号)T2Nov 3, 23:27·78 pts
ReSFlow Enables One-Step Generation for Discrete Data via Statistical Manifolds
- Achieves one-step generation for discrete data via spherical mapping and geodesic path planning
- On binary MNIST, a single model call matches the quality of 34 calls in Statistical Flow Matching
- Reduces conditional promoter DNA design time from 67.94s to 0.409s
View Related Reports & Analysis Original Source