Happy to introduce 🔥LaM-SLidE🔥! We show how trajectories of spatial dynamical systems can be modeled in latent space by --> leveraging IDENTIFIERS. 📚Paper: arxiv.org/abs/2502.12128 💻Code: github.com/ml-jku/LaM-SLidE 📝Blog: ml-jku.github.io/LaM-SLidE/ 1/n

We unlocked the working memory of LLMs 💥 Reasoning in Memory (RiM) replaces autoregressive "thinking out loud" with fixed memory blocks that form a task-specific workspace for latent reasoning. The key idea is simple: reasoning should happen inside the LLM, not in its output!

Symbol-equivariant Recurrent Reasoning Models (SE-RRM) SE-RRM advances HRM and TRM -- guaranteed identical solutions for problems with permuted colors (ARC AGI) or digits (Sudoku). Coolest part: extrapolation to larger problem sizes!!! P: arxiv.org/abs/2603.02193

# AI in Drug discovery just BROKE THROUGH a wall # A newer AI model, ConGLUDe, as fast but much more accurate than DrugCLIP. Instead on just 40K structure-based data, ConGLUDe is trained on 100M datapoints from ligand-based data P: arxiv.org/abs/2601.09693

The Great Comeback of Self-Normalizing Networks in 2025: It’s been a wild year in AI and for SNNs SELU!! See my overview and some trends here: Blog: bioinf-jku.github.io/SNNs/

📢 Another #NeurIPS, another diffusion circle! Join us to talk about diffusion models on Friday Dec 5 at 3:30PM in San Diego! Bayside terrace outside room 11 (upstairs) ☀️🚢🌊 Please help spread the word, tell your friends! No slides, no talks, we just sit down and chat 🗣️

Introducing our invited speakers at #ML4Molecules2025: Rocio Mercado ( rociomer.github.io ): Tenure-track professor at Chalmers University. Her work bridges machine learning and molecular discovery. Registration (free!): moleculediscovery.github.io/…

“In the judgement of the most competent living mathematicians, Fräulein Noether was the most significant creative mathematical genius thus far produced since the higher education of women began.” – Albert Einstein, 1935 (NYTimes) I’m thrilled to share that I’ll be joining emmi.ai, a company inspired by the legacy of Emmy Noether, for my upcoming internship. Over the next few months, I’ll have the opportunity to work with Sebastian Kaltenbach on my passion: Diffusion and Flow-based generative models, and their applications to physics. Excited for what lies ahead! While Noether devoted her life to uncovering the beauty of symmetries, our recent work explores a different path—approaching the problem without explicitly enforcing them. I’m proud that this work, done together with amazing collaborators @ArturToshev , Andreas Fürst, @gklambauer , @AndreasMayr11 , @jo_brandstetter , has been accepted to @NeurIPSConf 2025 in San Diego. arxiv.org/abs/2502.12128

Celebrating 4,000 citations! Thanks everyone who successfully used self-normalizing networks!!!

Nice, short post illustrating how simple text (discrete) diffusion can be. Diffusion (i.e. parallel, iterated denoising, top) is the pervasive generative paradigm in image/video, but autoregression (i.e. go left to right bottom) is the dominant paradigm in text. For audio I've seen a bit of both. A lot of diffusion papers look a bit dense but if you strip the mathematical formalism, you end up with simple baseline algorithms, e.g. something a lot closer to flow matching in continuous, or something like this in discrete. It's your vanilla transformer but with bi-directional attention, where you iteratively re-sample and re-mask all tokens in your "tokens canvas" based on a noise schedule until you get the final sample at the last step. (Bi-directional attention is a lot more powerful, and you get a lot stronger autoregressive language models if you train with it, unfortunately it makes training a lot more expensive because now you can't parallelize across sequence dim). So autoregression is doing an `.append(token)` to the tokens canvas while only attending backwards, while diffusion is refreshing the entire token canvas with a `.setitem(idx, token)` while attending bidirectionally. Human thought naively feels a bit more like autoregression but it's hard to say that there aren't more diffusion-like components in some latent space of thought. It feels quite possible that you can further interpolate between them, or generalize them further. And it's a component of the LLM stack that still feels a bit fungible. Now I must resist the urge to side quest into training nanochat with diffusion.

PAPER/ABSTRACT DEADLINE ALREADY END OF THIS WEEK! ELLIS Machine Learning for Molecules workshop: moleculediscovery.github.io/… DON'T MISS THE DEADLINE: short papers or extended abstracts welcome!

🚀 Excited to share our new paper on scaling laws for xLSTMs vs. Transformers. Key result: xLSTM models Pareto-dominate Transformers in cross-entropy loss. - At fixed FLOP budgets → xLSTMs perform better - At fixed validation loss → xLSTMs need fewer FLOPs 🧵 Details in thread

It's happening again!!! ML4Molecules workshop 2025. within the #ELLIS Unconference, preceding #EurIPS. More infos: moleculediscovery.github.io/…

Paper by bytedance, improves upon Meanflow by removing the need for JVP calculation

if you're interested in building the future of creative tools with us, we're hiring! krea.ai/careers

🚀 Introducing XBai o4：a milestone in our 4th-generation open-source technology based on parallel test time scaling！ In its medium mode, XBai o4 now fully outperforms OpenAI−o3−mini.📈 🔗Open-source weights: huggingface.co/MetaStoneTec/…✅ Github link: github.com/MetaStone-AI/XBai…

General relativity 🤝 neural fields This simulation of a black hole is coming from our neural networks 🚀 We introduce Einstein Fields, a compact NN representation for 4D numerical relativity. EinFields are designed to handle the tensorial properties of GR and its derivatives.

We release AB-UPT, a novel method to scale neural surrogates to CFD meshes beyond 100 million of mesh cells. AB-UPT is extensively tested on the largest publicly available datasets. 📄 arxiv.org/abs/2502.09692 🤗 huggingface.co/EmmiAI/AB-UPT 💻 github.com/Emmi-AI/AB-UPT

We introduce SIMSHIFT: A Benchmark for Adapting Neural Surrogates to Distribution Shifts. I sincerely hope that new ideas are coming out from this benchmark. Paper: arxiv.org/abs/2506.12007 Code: github.com/psetinek/simshift

Great discussion, @chaitjo! We also explored this with extensive experiments in our recent paper: arxiv.org/abs/2501.01999. We find, among others, that equiv mods in a sense scale even better than non-equiv ones. Going more or less completely against the vibes from your post😅1/5

A European-developed TiRex is leading the field—significantly ahead of U.S. competitors like Amazon, Datadog, Salesforce, and Google, as well as Chinese models from companies such as Alibaba.