Oscar Davis

You like discrete diffusion, but it's too slow? 🥀 You like test-time inference, but it's for continuous methods? 😩 We fixed it. Introducing Categorical Flow Maps: continuously sample discrete data in a single step 🚀💫 How? 🧵⬇️ 💪 Co-led with @FEijkelboom, @daan_roos_

"Categorical Flow Maps" by @daan_roos_ , @osclsd , @FEijkelboom , @mmbronstein, @wellingmax, @ismaililkanc , @LucaAmb , @jwvdm Paper: arxiv.org/abs/2602.12233 #machinelearning

Great concurrent work to our recent Categorical Flow Maps, which also show that it’s possible to generate high quality text in a single step. Their experiments further show how good it can get with more training. Check it out!

Can language models generate high quality full sequences in ONE step? Yes! Using continuous flows. We introduce Flow Map Language Models (FMLM) — a fundamentally new approach that outperforms discrete diffusion baselines across the board. 🔥SoTA few-step performance ❤️‍🔥 ≥8.3× faster sampling 🧵👇 “One-step Language Modeling via Continuous Denoising” (with amazing coauthors @wognsfjq96 , @mananag_007, Sheel Shah, @jrrhuang, @AdtRaghunathan, @hongseu33, @nmboffi, @jw9730)

🚨We present MacroGuide: the first model to generate arbitrary macrocycles in 3D 🎉 How? Topological guidance with persistent homology!🧵⬇️ [1/6] With @duplessis_alex @mmbronstein @AlexanderTong7 @fjduarteg @ismaililkanc

Really enjoyed working on this one with @osclsd and @FEijkelboom! Code for the graph experiments is now available: github.com/mrroose/catego… Working on a clean, minimal codebase to make it easy to build on — stay tuned.

Discrete Diffusion just got a huge upgrade! with "Categorical Flow Maps", it is now much faster + capable of test time inference. This paper shows you can make diffusion/flow models for discrete stuff (text, graphs, binary images) generate in 1–2 steps by flowing continuously on the probability simplex and self-distilling an endpoint-consistent “flow map”. This keeps predictions valid (on-simplex) and unlocks the same fast guidance tricks we use in continuous diffusion. So few-step speedups to categorical generation without the usual discrete-step pain!

Discrete diffusion — but fast? ⚡️ Test-time inference — but for discrete data? 🧠 Categorical Flow Maps: continuous transport toward the simplex, turning discrete generation into a single-step problem. Built on Variational FM (CatFlow), we obtain (self-)distillation from

[5/5] We think CFMs open a new path for discrete data generation. 💫 🔗 Paper: arxiv.org/abs/2602.12233 💾 Code: github.com/olsdavis/semic… Thanks to the incredible team! @FEijkelboom @daan_roos_ @mmbronstein @wellingmax @ismaililkanc @LucaAmb @jwvdm Thank you for reading! 🙂

[4/5] We also achieve SOTA results on molecular 🧪 and image generation 🎨 for the single step regime 🔥

You like discrete diffusion, but it's too slow? 🥀 You like test-time inference, but it's for continuous methods? 😩 We fixed it. Introducing Categorical Flow Maps: continuously sample discrete data in a single step 🚀💫 How? 🧵⬇️ 💪 Co-led with @FEijkelboom, @daan_roos_

RT @bose_joey: A new joint Phd opportunity @imperialcollege with through the School of Convergence. The Phd is truly interdisciplenary and…

The last paper of my PhD got into #ICLR2026 !! 🚀 ❓Why do diffusion models generalise instead of memorising training data? 💡 Implicit regularisation! We take a learning theoretic approach to identifying mechanisms fundamental to their training/sampling arxiv.org/abs/2507.03756

@pranamanam Haha, thanks, Pranam! 😉

@osclsd Let's goo! This is so beautiful. :)

A bit late to the party, but glad to say we'll see you in Rio :) 🇧🇷 #ICLR2026

@msalbergo Thanks, Mike! :)

@osclsd Congrats Oscar!

@valentin_kilian @OxfordYss Le patron en plein milieu

So glad to be part of this group! A big thank you to G-Research for making it possible for me to travel to San Diego to present my work and support the rest of the @OxfordYss at #NeurIPS2025.

We're presenting “Amortized Sampling with Transferable Normalizing Flows” today at NeurIPS! Find us at Poster #1614 (11am–2pm, Exhibit Hall C/D/E) 🔍 Let’s chat sampling, molecular dynamics, and normalizing flows! 😊