Fantastic work scaling up neural network VMC methods like FermiNet and Psiformer to previously impossible systems - you can start to tackle real challenging chemistry at this scale!

Happy to share that our paper "Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space" got accepted to #ICLR2025, and we will be presenting it this week in Singapore! Joint work with Nicholas Gao, Tom Wollschläger, Johanna Sommer, and Stephan Günnemann. 🧵 1/

🎉Excited to announce our #ICLR2025 Spotlight! 🚀Lukas Gosch and I will be presenting our paper on the first exact certificate against label poisoning for neural nets and graph neural nets. Joint work with Stephan Günnemann and Debarghya 👇[1/6]

Thrilled to announce that we just presented „MAGNet: Motif-Agnostic Generation of Molecules from Scaffolds“ at #ICLR2025 🧲 Johanna Sommer Bastian Grossenbacher-Rieck Fabian Theis Stephan Günnemann For those who couldn’t make it to our spotlight: openreview.net/forum?id=5FXKg…

If you are attending #ICLR2025 and are interested in electronic structure modelling / quantum chemistry come by our poster on learnable non-local XC-functionals to discuss with Nicholas Gao and me. 🗓️ Today | 3:00 pm– 5:30 pm 📍Hall 3 | Poster #3

How do LLMs navigate refusal? Our new @ICMLConf paper introduces a gradient-based approach & Representational Independence to map this complex internal geometry. 🚨 New Research Thread! 🚨 The Geometry of Refusal in Large Language Models By Stephan Günnemann's lab & Google AI. 🧵👇

I am very happy to share Orbformer, a foundation model for wavefunctions using deep QMC that offers a route to tackle strongly correlated quantum states! arxiv.org/abs/2506.19960

For a short ML-focused introduction to this paper, I wrote a short blog post ae-foster.github.io/posts/2025/06/… (3/n)

Describing strongly correlated quantum systems remains a major challenge in quantum chemistry. Deep QMC offer a potential solution, but at a huge computational cost. (4/n)

To get cost down we make use of amortization: solving a single minimization problem with a more complex network that represents multiple wavefunctions simultaneously (5/n)

We scaled this idea up and pushed it to work on strongly correlated systems. On a cost/error plot, we find that Orbformer is on or ahead of the Pareto frontier formed by traditional multireference methods, a first for deep QMC. (6/n)

We also saw very strong results confirming the experimental activation energy of a Diels-Alder reaction, and significantly outperforming earlier transferable QMC approaches (7/n)

Digging into the model we found intriguing behaviour, such as the unsupervised discovery by the model of ‘core’ electron orbitals for second row atoms. This has been a fascinating project to be a part of! Check out the preprint for more details and results. (8/n)

How private is DP-SGD for self-supervised training on sequences? Our #ICML2025 spotlight shows that it can be very private—if you parameterize it right! 📜arxiv.org/abs/2502.02410 #icml Joint work w/ M. Dalirrooyfard, J. Guzelkabaagac, A. Schneider, Y. Nevmyvaka, Stephan Günnemann 1/6

Real data is noisy but HiPPO assumes it's clean. Our UnHiPPO initialization resists noise with implicit Kalman filtering and makes SSMs robust without architecture changes. #ICML poster: Thu 11am E-2409 Paper: openreview.net/forum?id=U8GUm… Code: github.com/martenlienen/u… w/ Stephan Günnemann

Want to learn more about how to make self-supervised training on sequential data differentially private? Come by our #ICML2025 poster E-1007 in East Exhibition Hall A/B on Tuesday at 11am! #icml

Injecting structure into LLMs with no changes to the architecture? SAFT🧃 Structure-aware LLM fine-tuning for AMR-to-text. New SOTA + no model changes! 📍#SKnowLLM, KDD, Toronto 🇨🇦 🗓️ Aug 4, 1–5pm · Room 717 📄 arxiv.org/abs/2507.13381 w/ R. Kamel, Simon Geisler, Stephan Günnemann 1/5