Skip to main content

Categorical Flow Maps

Generating categorical data in one or a few steps with continuous flow maps and self-distillation.
Author

Oscar Davis (Oxford)

Published

19 November 2026

Time: 14:00–15:00 (Munich time; Europe/Berlin)

Location: European Southern Observatory (ESO), Phoenix (C.2.02).

Paper summary

Categorical Flow Maps aim to speed up the generation of discrete data such as text and molecular graphs. The method uses continuous trajectories with endpoint predictions on a probability simplex, allowing a model to learn larger jumps along the generative path. Self-distillation reduces sampling to one or a few steps, while the continuous formulation supports guidance toward chosen objectives. The paper evaluates the approach on molecular graphs, binary images and text.

Categorical Flow Maps overview comparing velocity sampling, flow-map sampling, and Lagrangian training, alongside generated molecule and text examples.

Figure 1 from Roos et al., Categorical Flow Maps, reproduced unchanged under CC BY 4.0. Select the image to view it at full size.

Materials

Paper on arXiv (2602.12233)

Slides (PDF)