(I/III) We're excited to announce a new tenure track opening! The position is called 'quantum informatics' and is affiliated with our QUICK group within the CS+AI division at @jku.at 🇦🇹. Application deadline is November 30th, 2025: www.jku.at/en/the-jku/w...

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

Deadline extended to October 15th!

PAPER/ABSTRACT DEADLINE ALREADY END OF THIS WEEK! ELLIS Machine Learning for Molecules workshop: moleculediscovery.github.io/workshop2025/ DON'T MISS THE DEADLINE: short papers or extended abstracts welcome!

ELLIS ML4Molecules workshop 2025 ** Call for papers** is out: moleculediscovery.github.io/workshop2025/

Is Bluesky the actual dead Internet??

It's happening again!!! ML4Molecules workshop 2025. within the #ELLIS Unconference, preceding #EurIPS. More infos: moleculediscovery.github.io/workshop2025/

Happy that IBM uses proper MLPs, i.e. self-normalizing networks, in their Flowstate model! ;) P: arxiv.org/abs/2508.05287

Another win for xLSTM

It just struck me that the ancient Greeks chose their twelve Olympian gods with gender parity..

We Austrians are fully aware that we are a democratic country now, but we still sneak in pictures of our monarchs... :)

Controversial posts are downranked on Blue sky???? Oh c'mon...

Machine Learning-Driven Optimization of Specific, Compact, and Efficient Base Editors via Single-Round Diversification New base editors designed and optimized with ML! Almost all ML-designed ones active in the lab tests!! P: www.biorxiv.org/content/10.1...

Devastating blow for GNNs/MPNNs for chemistry and small molecules: All of them are not statistically better than simple ECFP . Only our CLAMP model does the job ;) P: arxiv.org/abs/2508.06199

Just in: GPT-5 achieves AGI in chemical sciences!!1!!1! Actually, NO! .... struggles with generating 2D graphs from SMILES.

Great idea for a dataset and task! Thanks for using and assessing out Prot-xLSTM

We are soooo proud. Our 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.

Ever though you could get a 35 million parameter time series state-of-the art foundation model that you can run on embedded hardware? Thanks to @hochreitersepp.bsky.social and his team at NXAI, you can. Amazing work! Paper: arxiv.org/abs/2505.23719 Code: github.com/NX-AI/tirex

Recommended read for the weekend: Sepp Hochreiter's book on AI! Lots of fun anecdotes and easily accessible basics on AI! www.beneventopublishing.com/ecowing/prod...

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-S... 📝Blog: ml-jku.github.io/LaM-SLidE/ 1/n

Latent Space Modeling of Spatial Dynamical Systems via Linked Entities LATENT DIFFUSION and GRAPH NEURAL NETWORKS finally combined!! Best model for molecular dynamics of small molecules!!! Paper: arxiv.org/abs/2502.121... Blog: ml-jku.github.io/LaM-SLidE/

Join the tutorial on xLSTM!!

Talking at CAIML Symposium today "A New Generation of Foundation Models Based on xLSTM" Slides available here: cloud.ml.jku.at/s/fLtjm5GKQS... caiml.org/news/194/

Transformer-dot-com bubble.

In-context learning for molecules. HuggingFace App. Try it out!!!

TENURE TRACK PROFESSORSHIP in Reinforcement Learning announced at Johannes Kepler University Linz. Three other ML/AI tenure tracks also open. www.jku.at/en/the-jku/w...

Talking about AI systems and their slow uptake into healthcare systems and clinical practice today at dHealth: dHealth 2025 conference: dhealth.at/programme/ Slides: cloud.ml.jku.at/s/sNs3Z2GD2C...

Leveraging Partial SMILES Validation Scheme for Enhanced Drug Design in Reinforcement Learning Frameworks Uncomplete SMILES strings already assessed during autoregressive generation. Decreases amount of incorrect molecules. P: arxiv.org/abs/2505.00530

May the 4th be with you!

MHNfs: Prompting In-Context Bioactivity Predictions for Low-Data Drug Discovery Few-shot models for molecules now easily accessible via web application. Predictions via prompting SMILES P: pubs.acs.org/doi/10.1021/...

Softpick: No Attention Sink, No Massive Activations with Rectified Softmax Work claims that asing softpick instead of softmax in LLM training removes attention sinks and massive activations. Run with 0.37B parameter LLM similar to normal attention P: arxiv.org/abs/2504.20966

Come to Austria and work in AI for Life Sciences!!

Knowledge Distillation for Molecular Property Prediction: A Scalability Analysis Distillation does not only work for LLMs, but also for activity/property prediction models for molecules. P: advanced.onlinelibrary.wiley.com/doi/full/10....

Augmented and Programmatically Optimized LLM Prompts Reduce Chemical Hallucinations Topological polar surface area (TPSA) predicted by an LLM. Fewer confabulations with a RAG systems plus some additional safeguards. P: doi.org/10.1021/acs....

1/11 Excited to present our latest work "Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics" at #ICLR2025 on Fri 25 Apr at 10 am! #CombinatorialOptimization #StatisticalPhysics #DiffusionModels

Unified Molecule Generation and Property Prediction Transformer trained to jointly generate molecules and predict molecular properties. Task-tokens used for different properties. Encoder-decoder architecture. P: arxiv.org/abs/2504.16559

Diverse and feasible retrosynthesis using GFlowNetsDiverse and feasible retrosynthesis using GFlowNets Predicting/generating chemical reactions. Semi-template approach with GFlowNets as generative model. Lots of evaluations, e.g. multi-step retrosynthesis. P: www.sciencedirect.com/science/arti...

A User-Tunable Machine Learning Framework for Step-Wise Synthesis Planning Chemical reactions & synthesis with NEURO-SYMBOLIC AI: - (modern) Hopfield networks used - Reaction templates (symbolic rules) Includes tunable scores for cost, temp, and activity. 🧪🧠 P: arxiv.org/abs/2504.02191

Huge achievement: arxiv.org/abs/2503.14376 xLSTM kernels are now the fastest kernels both for training and inference. Faster than Flashattention or Mamba kernels. High arithmetic intensity. Optimized use of GPUs that set a new state of the art. Congratulations to the team. It is highly impressive.

xLSTM-based FOUNDATION MODELS FOR LIFE SCIENCES Talk at Helmholtz-ELLIS Workshop on Foundation Models in Science today. Slide deck available here: cloud.ml.jku.at/s/trjeiy5NQW... Workshop webpage: ellis.eu/events/helmh...

I agree, and very much related I want to promote my old Lab, led my @hochreitersepp.bsky.social and including researcher like @gklambauer.bsky.social and @jobrandstetter.bsky.social. They have 5 (!) open PostDoc position and 10 (!!) open PhD positions. See bsky.app/profile/hoch...

xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference Meet the fastest 7B language model out there. Based on the mLSTM! P: arxiv.org/abs/2503.13427

The new #RDKit blog post is an updated tutorial on constructing simple SAR tables in Jupyter notebooks: greglandrum.github.io/rdkit-blog/p...

Opportunity to work with @hochreitersepp.bsky.social , @jobrandstetter.bsky.social , and me!! We have many open positions in machine learning, deep learning, LLMs!! Both for PostDocs and PhDs! Join us!

xLSTM for Automated Stock Trading: arxiv.org/abs/2503.09655 xLSTM outperforms LSTM. "These findings mark the potential of xLSTM for enhancing DRL-based stock trading systems." This reinforcement learning approach uses xLSTM in both actor and critic components, which increases the performance.

NORMALIZER-FREE DEEP NETS 2017: normalizer-free deep MLPs arxiv.org/abs/1706.02515 2021: normalizer free ResNets arxiv.org/abs/2102.06171 2025: normalizer-free Transformers arxiv.org/abs/2503.10622 by @yann-lecun.bsky.social and colleagues Normalizer-free xLSTM in 2029?

Happy Pi day!!!

The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems A set of tasks/question for LLMs where they are pressured to lie (e.g. to "costumers"). Grok has highest tendency to lie; Claude 3.7 least.. P: arxiv.org/abs/2503.03750

Another Turing Award for machine learning!!! Wooow