Associate Professor (UHD) at the University of Amsterdam.
Amsterdam, NL
Joined March 2009
- Tweets1.1K
- Following1.4K
- Followers2.7K
- Likes3.3K
Jan-Willem van de Meent retweeted
Continuous Diffusion LMs went from niche to contender this year, but the recipe is shaky: why self-conditioning? How to embed tokens? 🤨
Our new work takes your one-hot flow from 4% to 50% on TinyGSM without self-conditioning, and explores the design space of CDLMs. 🔥🚀 (1/6)
Jan-Willem van de Meent retweeted
And the harder the problem, the bigger the gap: larger mazes, more objects per CLEVR scene (85% vs 62% at 10 objects).
What diffusion lacks on rule-governed data isn't capacity. It's a place to think.
📄 arxiv.org/abs/2610.09876
w/ Paweł Skierś, @magmatic624 @e7mul @jwvdm
Jan-Willem van de Meent retweeted
Sad to be missing #COLM2026 but check out Particles of Thought at the Efficient Reasoning workshop @COLM_conf !
w/ @wellingmax & @jwvdm, we align process supervision with test-time search by applying Sequential Monte Carlo at both training and inference.
openreview.net/forum?id=zBqC…
Jan-Willem van de Meent retweeted
[1/9 🧵]
♾️ Diffusion Transformers can scale in two ways: by making the network larger, or by evaluating it more times during sampling (= more integration steps).
Looped transformers suggest a third possibility: reuse the same parameters repeatedly, increasing executed depth without increasing model size.
Jan-Willem van de Meent retweeted
Now I think I understand how policy professionals who spent their careers trying to cut the Gordian knot of American healthcare must have felt when Bernie waddled in and just shouted "we're just gonna do Medicare for all!"
Jan-Willem van de Meent retweeted
I have to assume everyone who believed his campaign line about "stapling a visa to every diploma" was just in some kind of medically induced coma between 2017 and 2021.
This fee would be by far the biggest move to restrict high-skilled immigration to date.
Some may bend over backwards to convince themselves that this would be some moderate compromise short of ending OPT. In reality, this fee would nuke America's largest and most important program for new high-skilled immigrants.
The international student pipeline is how the vast majority of top talent first comes to the United States. Per CSET, 72 percent of immigrant founders of top AI companies first came to the country on student visas. OPT is the primary way through which foreign-born employees at places like Tesla, Intel, NVIDIA, Open AI, DeepMind, Anthropic, etc. enter the US labor market.
Without OPT, the rest of the immigration system would become less skilled and lower-paid. More H-1B slots would go to IT outsourcers and staffing companies who use the system for cheap labor in mid-skilled jobs. This effect would grow over time, as this fee would deter international students from coming here in the first place.
The costs of such a rule, if sustained, to US innovation, entrepreneurship, and economic growth would be massive. If you've been on the sidelines waiting for the shoe to drop on high-skilled immigration, this is it.
Jan-Willem van de Meent retweeted
Any indication on we might get the #ICLR2027 @iclr_conf CfP or submission deadline? Asking for a friend...
Jan-Willem van de Meent retweeted
It’s showtime at #ICML2026! 🚀 Really excited to be presenting our paper, "Platonic Transformers," today. We’ve developed a novel equivariant architecture that is finally, properly scalable. Come find me, @Mohammad_Niaz94 and @SharvVadgama at Hall A poster #2614!
Jan-Willem van de Meent retweeted
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. 🧵👇
Jan-Willem van de Meent retweeted
Every step of working on this I was amazed it actually works. The ways of deep learning are truly beautiful.
Amazing work led by Pedro, who is currently finishing his master thesis at AMLab.
Joint work with @FEijkelboom and @jwvdm.
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. 🧵👇
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!
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. 🧵👇
Jan-Willem van de Meent retweeted
🏎️Drift in the right direction🏎️
Introducing kernel-gradient drifting models: a reformulation of drifting models where the kernel itself defines the direction of motion through its gradient.
📜Paper: arxiv.org/pdf/2605.10727
💾Notebook: tinyurl.com/mv2jhuky
Jan-Willem van de Meent retweeted
Very excited about our work on finding the right drifting direction 🐎
We tackle a core open question in drifting: when does “no drift left” mean the model really matched the data?
Kernel-gradient drifting is the answer (with natural extensions to manifolds + discrete data)!
🏎️Drift in the right direction🏎️
Introducing kernel-gradient drifting models: a reformulation of drifting models where the kernel itself defines the direction of motion through its gradient.
📜Paper: arxiv.org/pdf/2605.10727
💾Notebook: tinyurl.com/mv2jhuky
Jan-Willem van de Meent retweeted
🚀 Excited to share my @GoogleDeepMind student researcher project: Dual-Rate Diffusion✨
⚡ A simple construction that speeds up both regular diffusion and distilled models by interleaving a heavy context encoder with a light conditional denoiser.
🧵👇
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.
Jan-Willem van de Meent retweeted
🌍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!
Jan-Willem van de Meent retweeted
We were all wondering whether Categorical Flow Maps (CFMs) could scale... 🤔
I couldn't help trying it out...
So we scaled CFMs to 1.7B parameters over 2.1T tokens 🚀🔥
Short summary 🧵⬇️
Jan-Willem van de Meent retweeted
In case you missed it — we're hiring an Assistant Prof in Machine Learning (AI4Science) at @AmlabUva
🇪🇺 One of Europe's top ML groups, based in Amsterdam.
Deadline: May 30 — still time to apply! 👇
Excited to share that we're looking for a new colleague at @AMLab_UvA : Assistant Professor in AI for Science 🔬🤖
AMLab is a world-class ML research group embedded in Amsterdam's thriving AI ecosystem: leading research groups, an ELLIS unit, startups, and big tech — all within reach.
And Dutch academic labor conditions are genuinely among the best in Europe ❤️
Deadline: May 30 👉 werkenbij.uva.nl/en/vacancie…
Jan-Willem van de Meent retweeted
Categorical Flow Maps got accepted to #ICML2026 🎉
See you in 🇰🇷 :D
Many thanks to @osclsd @daan_roos_ @mmbronstein @wellingmax @ismaililkanc @LucaAmb and @jwvdm ♥️
Wanna know more about flow-based LLMs: flow-based-llms.github.io/