Pinned Tweet
Can a lightweight Transformer compete in crystal generation without equivariance?
We show that it can.
Crystalite combines chemistry-aware priors with geometric attention biases for:
- SOTA CSP
- best S.U.N. among all baselines
- much faster sampling
More info below! 👇
Tin Hadzi Veljkovic retweeted
Our survey on continuous diffusion and flow language models is out! ❤️
A little animated tour of what’s inside 👇 From representations and training to few-step generation, scaling, and open problems.
continuous-lm.github.io
New survey on Continuous Diffusion and Flow Language Models is out! 🌊💬
One of my first papers, back in early 2025, was on continuous simplex flows. After working on discrete diffusion, I wanted to try controllable sequence generation with flow matching, which naturally operates in continuous spaces. But at the time, I found them much harder to train well than discrete diffusion. 🥲
2026 has changed this picture completely! 🚀 It’s been super exciting to watch continuous LMs take off, but with the rapid advancements, it’s also becoming harder to keep up with all the new work across pre-training, few-step generation, guidance, and scaling 🥹
So together with @chandavidlee, @FEijkelboom, and @jw9730, we put together a survey that consolidates the continuous LM literature and its key insights into 37 pages. It covers everything from the early work before the so-called “winter” of continuous LMs to the wave of new papers in 2026 🌟
More in thread! 🧵
Tin Hadzi Veljkovic retweeted
Excited to share that our paper, “Kernel-Gradient Drifting Models,” has been accepted to NeurIPS 2026! ❤️
📄 Paper: arxiv.org/pdf/2605.10727v2
💻 Code: github.com/jcarrascopollo/kg…
Tin Hadzi Veljkovic 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!
Tin Hadzi Veljkovic retweeted
Woo! #ICLR2026 Diffusion circle is happening tomorrow :D
We’ll meet above the registration desk - can’t wait to meet everyone ❤️
⌚️ Sat, Apr 25, 2:45
📍 Right above the registration desk and below the oral sessions
Who's at #ICLR2026 next week?
I'm not (😢), which means someone else is going to have to organise a diffusion circle! I'm counting on you!
Let me know, happy to amplify the announcement📢
Tin Hadzi Veljkovic retweeted
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 :)
Tin Hadzi Veljkovic retweeted
✨Excited to share that our work “MSPT: Multi-Scale Patch Transformer for Efficient Large-Scale Physical Modeling” was selected as a #CVPR2026 Highlight!
📰 Paper: arxiv.org/abs/2512.01738
🤗 Project page: pedrocurvo.com/projects/mspt
💻 Code: github.com/pedrocurvo/mspt
🏆State-of-the-art performance on PDE benchmarks (elasticity, fluids, porous flow) and large-scale CFD tasks like ShapeNet-Car.
---
Scaling neural PDE solvers to millions of spatial elements requires capturing two things at once:
1⃣ fine-grained local interactions: stress concentrations, small-scale turbulence, sharp gradients;
2⃣ long-range global dependencies: pressure coupling in fluids, far-field boundary conditions in aerodynamics;
Current methods tend to trade one for the other: local attention preserves detail but misses global structure, while global aggregation (e.g., fixed-size slices) captures the big picture but oversmooths fine scales.
MSPT uses Parallelized Multi-Scale Attention (PMSA) to capture local and global interactions within a single, efficient attention call.
Key highlights:
📈Near-linear scaling to millions of points on a single GPU
🔎Unified local-global modeling: fine-grained interactions within patches + global context through shared supernodes.
🚀 Fast and expressive: pushes the Pareto frontier on PDE modelling.
🌳Flexible across geometries: supports arbitrary meshes and point clouds via ball-tree partitioning.
Congrats to @pmpcurvo who pulled it off, co-supervised with @jwvdm.
Tin Hadzi Veljkovic retweeted
Sorted! @FEijkelboom and @GrigoryBartosh will be your #ICLR2026 diffusion circle organisers 🙏 stay tuned!
Who's at #ICLR2026 next week?
I'm not (😢), which means someone else is going to have to organise a diffusion circle! I'm counting on you!
Let me know, happy to amplify the announcement📢
AMLab is excited to present 7 papers at ICLR 2026! See you all in Rio 🇧🇷
1. iclr.cc/virtual/2026/poster/…
2. iclr.cc/virtual/2026/poster/…
3. iclr.cc/virtual/2026/poster/…
4. iclr.cc/virtual/2026/poster/…
5. iclr.cc/virtual/2026/poster/…
6. iclr.cc/virtual/2026/poster/…
7. iclr.cc/virtual/2026/poster/…
Links here:
📄 Paper
arxiv.org/abs/2604.02270v1
📝 Blogpost
stases.github.io/blogposts/c…
💻 Code
github.com/joshrosie/crystal…
Feel free to reach out if this is interesting or relevant to your work!
1
9
878
Tin Hadzi Veljkovic 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_
Tin Hadzi Veljkovic retweeted
Dynamic Training Enhances Machine Learning Potentials for Long-Lasting Molecular Dynamics #MolecularDynamics
pubs.acs.org/doi/10.1021/acs…
@HvTin
#JCIM Vol65 Issue15 #MachineLearning #DeepLearning
Tin Hadzi Veljkovic retweeted
Flow Matching (FM) is one of the hottest ideas in generative AI - and it’s everywhere at #ICML2025.
But what is it? And why is it so elegant? 🤔
This thread is an animated, intuitive intro into (Variational) Flow Matching - no dense math required.
Let's dive in! 🧵👇
Tin Hadzi Veljkovic retweeted
Excited to share Rotating Features – accepted as oral at #NeurIPS2023
Rotating Features learn to represent object affiliation via their orientation on real-world data without labels.
Let’s dive in!
📜 arxiv.org/abs/2306.00600
🖥️ github.com/loeweX/RotatingFe…
Tin Hadzi Veljkovic retweeted
Announcing Best Paper Awards for ML Reproducibility Challenge 2021!
We had over 100+ submissions and we are happy to accept 43 reports in our main program. Congratulations to our best and outstanding paper award winners. See more here: paperswithcode.com/rc2021
Tin Hadzi Veljkovic retweeted
Allright! First lecture is done 🥳 I'll be filling up this YT playlist with more videos on *group equivariant deep learning* in the coming weeks.
Regular g-convs: Done! Next up: Steerable g-convs!
youtube.com/playlist?list=PL…























