
Machine Learning PhD student at @AmlabUva
Amsterdam, Netherlands
Joined July 2011
- Tweets51
- Following429
- Followers213
- Likes3K
Super fun project that shows that even (masked) discrete dLLMs can generate multiple tokens in-parallel just fine!
Diffusion LLMs are meant to be fast ⚡️: they can decode many tokens in parallel in a single step.
The catch: push the parallelism too far and generation quality drops sharply.
We introduce a simple fix: Temporal Self-Distillation (TSD): the model learns from its own later, more-confident predictions, using only the rollout a verifier marks correct 👇🧵
Metod Jazbec retweeted
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…
Metod Jazbec retweeted
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 scratch.
Language, molecules, and test-time steering — one framework.
Scaling to LLMs and foundation models next. Watch this space 👀
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_
Presenting today at @NeurIPSConf from 11am-2pm (poster number #3904)👇👇
TTA Monitor is coming to #NeurIPS2025!
Our monitor flags risk increase of a continuously adapting model without any labels and with false-alarm guarantees.
Let’s dive in! 🧵 1/6
📄 Paper: arxiv.org/abs/2507.08721
📌 Poster: neurips.cc/virtual/2025/loc/…
This was a super fun project with @monaschir, see you all at NeurIPS soon! ✈️
TTA Monitor is coming to #NeurIPS2025!
Our monitor flags risk increase of a continuously adapting model without any labels and with false-alarm guarantees.
Let’s dive in! 🧵 1/6
📄 Paper: arxiv.org/abs/2507.08721
📌 Poster: neurips.cc/virtual/2025/loc/…
2
176
Metod Jazbec retweeted
Come by our Oral on Monitoring Risks in Test-time Adaptation at the @PUT_TTA_ICML25 #ICML Workshop
⏱️: Friday, 10am PDT
🚩: West Meeting Room 220-222
Work with @MetodJazbec @canaesseth @eric_nalisnick
Metod Jazbec retweeted
📜 Monitoring Risks in Test-Time Adaptation
(ICML PUT Workshop Oral!)
Time: Fri 18 Jul 10 a.m. PDT
Location: West Meeting Room 220-222
Presenter: @monaschir
arxiv.org/abs/2507.08721
Great too see that our Generative Uncertainty won the best paper award at the ICLR QUESTION workshop (uncertainty-foundation-model…). If you're interested in what Bayesian/ensembling methods can bring to the world of diffusion models, check out the paper 👇
arxiv.org/abs/2502.20946
🥇 3 Best Paper Awards for AMLab members!
1. Towards Variational Flow Matching on General Geometries by @olgazaghen @FEijkelboom
2. Generative Uncertainty in Diffusion Models by @MetodJazbec
3. SDE Matching by @GrigoryBartosh
Congrats to everyone! 🔥
Metod Jazbec retweeted
Generative Uncertainty in Diffusion Models (spotlight)
by @MetodJazbec Eliot Wong-Toi, Guoxuan Xia, Dan Zhang, Eric Nalisnick, Stephan Mandt
➡️ openreview.net/forum?id=K54V…
⚠️ Quantify Uncertainty and Hallucination in Foundation Models: The Next Frontier in Reliable AI
Come say hi to me, @AlexTimans, and @eric_nalisnick today between 11am-2pm in East Exhibit Hall A-C #4505 . We'll be presenting our Fast yet Safe paper on efficient inference and risk control 👇
Hey! Catch me next week at @NeurIPSConf where we’ll be presenting our Fast yet Safe paper. We show how to equip efficient AI methods like early-exiting or soft speculative decoding with performance guarantees using risk control!
arxiv.org/abs/2405.20915
🧵🔽
I will also be presenting 3 workshop papers at @NeurIPSConf , all related to the efficiency of LLMs/diffusion
🧵👇
1
11
359
And finally in “On Efficient Distillation from LLMs to SLMs”, we investigate the importance of incorporating student model’s feedback for data efficiency when generating synthetic data for knowledge distillation between LLMs and SLMs
openreview.net/attachment?id…
1
2
123
All there projects are a result of very fun collaborations with Master students at @UvA_Amsterdam and my summer internship at @Microsoft
If you also find efficiency and robustness of LLMs/diffusion interesting, do reach out next week, happy to chat more about it!
3
120
Hey! Catch me next week at @NeurIPSConf where we’ll be presenting our Fast yet Safe paper. We show how to equip efficient AI methods like early-exiting or soft speculative decoding with performance guarantees using risk control!
arxiv.org/abs/2405.20915
🧵🔽
In our experiments, we apply our approach to calibrate a broad set of models for different tasks, ranging from machine translation using LLMs to image generation with diffusion
1
1
110
If this caught your attention, have a look at out paper and come talk to us during the NeurIPS poster session on Thu 12 Dec at 11:00 a.m. PST (East Exhibit Hall A-C #4505)
Great collaboration with: @AlexTimans, @eric_nalisnick, @canaesseth, @isDanZhang, @HvTin and others
















