Twitxy Saved

Jan-Willem van de Meent

@jwvdm

Associate Professor (UHD) at the University of Amsterdam.
Amsterdam, NL jwvdm.github.io Joined March 2009
1.4K Following 2.7K Followers 1.1K Posts

Latest posts 5

Jan-Willem van de Meent @jwvdm
Incredibly proud of Floor, who was selected for the Google PhD fellowship this year. Very well deserved!
Floor Eijkelboom Verified account @FEijkelboom 5d
Super excited to have been selected as a 2026 Google PhD Fellow for my work on flow-based language modeling! 🌊

Big thanks to my advisors @jwvdm and @wellingmax, everyone I've worked with, and @Google for the support! ❤️

blog.google/company-news/o…
View quoted post on X
1 replies 0 reposts 36 likes 3.4K views View on X
Jan-Willem van de Meent @jwvdm
@pmpcurvo and @mmderakhshani show that flow matching with looped transformers pays off: 60% fewer parameters, 52% fewer inference FLOPs and a 3.3 lower FID than DiT-XL/2. Really great work!
Pedro Verified account @pmpcurvo Oct 6
♾️ LiFT: Loop Flow Transformers 🧵👇

Diffusion models usually spend more inference compute by taking more sampling steps. What if we could also spend compute inside each step?

LiFT turns part of a DiT into a recurrent core, letting us scale depth at inference with only a small architectural change. On ImageNet, it beats dense DiT-XL/2 with 3.34 lower FID, while using 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.

Jointly led with Mohammad Mahdi Derakhshani 👑(@mmderakhshani)

And amazing supervisors Cees Snoek (@cgmsnoek), Jan-Willem van de Meent (@jwvdm) and Gertjan J. Burghouts (@gjburghouts)
View quoted post on X
1 replies 2 reposts 26 likes 1.6K views View on X
Jan-Willem van de Meent @jwvdm
Super excited about this new paper with Pedro Curvo. To guide a flow model at sampling time, just steer toward the empirical mean of a set of reference samples. It's almost embarrassingly simple, and it works much better than it has any right to. Worth a read!
Pedro Verified account @pmpcurvo May 21
Guide with examples, not rewards 🐘

Controlling what a pretrained generative model produces is still mostly a choice between three slow options: fine-tune it, attach a reward network, or search at inference. We found flow matching allows a fourth, and it costs almost nothing.

In deterministic interpolants, the velocity of the flow is determined by where the trajectory is headed: the endpoint mean. Shift that mean, and the entire flow shifts with it.

This turns control into a matter of reference. Change the examples that define the endpoint, and you change the direction the model follows. The examples need not be perfect. They only need to point the flow toward the attribute you want.
Color, identity, style, and structure, all controllable through examples. 🧵👇
View quoted post on X
0 replies 6 reposts 37 likes 3.8K views View on X
Jan-Willem van de Meent @jwvdm
Great paper by Maksim, who has been building up a repertoire of signature scalable attention mechanisms for physical prediction tasks! Mosaic produces sharp, calibrated weather ensembles; global ensemble forecasts in seconds on a single H100.
Max Zhdanov Verified account @maxxxzdn May 20
🌍Today we release Mosaic, a probabilistic weather model that shifts the Pareto frontier of ML weather forecasting.

It matches the skill of state-of-the-art models while generating a 24-member, 10-day global forecast in under 12 s on a single H100.

Thread!
View quoted post on X
0 replies 5 reposts 13 likes 2.6K views View on X
Jan-Willem van de Meent @jwvdm
Floor (@FEijkelboom) wrote a crisp and clear introduction to variational flow matching for text generation. He also discusses recent extensions to learning flow maps (arxiv.org/abs/2602.12233). Really nice animations throughout!
arxiv.org Categorical Flow Maps We introduce Categorical Flow Maps, a flow-matching method for accelerated few-step generation of categorical data via self-distillation. Building on recent variational formulations of flow...
Floor Eijkelboom Verified account @FEijkelboom Apr 22
Flow-LLM Blogpost :D flow-based-llms.github.io

In the last few weeks, a bunch of work on flows for language came out 🌊

That is exciting, because it makes truly parallel text generation feel real: generation where models can keep refining the whole response during inference, instead of committing token by token.

I wrote an intuitive and animated introduction to the area — why autoregression has a structural ceiling, why discrete diffusion only partly escapes it, and why flows may be the first genuinely parallel alternative.

Here's an overview of the key parts of the blog - and let's chat at #ICLR2026 :)
View quoted post on X