Pinned Tweet
New preprint:
"Transferable SCF-Acceleration through Solver-Aligned Initialization Learning (SAIL)"
* Standard ML initial guesses target the converged density.
* We target the solver trajectory instead.
Eike Eberhard retweeted
Long overdue news: I have since moved to London 🇬🇧 and started my PhD at @imperialcollege with @bose_joey, working on generative models for science.
If you are in London and this is your kind of thing, I'd be glad to meet and talk about generative models for bio and beyond ☕
Eike Eberhard retweeted
Why should a 3D molecule generator learn a benzene ring atom by atom? 🤔
MotiFlow generates molecules from rigid motifs instead: one joint flow over motif types and their SE(3) poses.
#NeurIPS2026 🎉 with @MarcelKollovieh, @ESEberhard, and @guennemann 🧵
Eike Eberhard retweeted
Excited to be presenting "Diffusion LLMs are Natural Adversaries for any LLM" at #NeurIPS2026 in Sydney 🐨
Come by poster session 4, Dec 9 5pm
With @PaulUngermann,@TomWollschlager, @guennemann,@SchwinnLeo
PS: finishing my PhD, looking for research roles in generative models.
Open-source Diffusion LLMs easily break GPT-5!
In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access.
@TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn
🧵
Eike Eberhard retweeted
Discrete flow matching usually changes one component per edit in a fixed-size state. Can we extend it to general actions, like inserting whole molecular fragments?
Introducing Trajectory Matching 🧵
New paper with M. Lienen, @ludke_david, @guennemann
Eike Eberhard retweeted
Blog Post Track is happening again at ICLR this year 🥳
Looking forward to your submissions!
(Please no slop)
See: iclr-blogposts.github.io/202…
@iclr_conf
Eike Eberhard retweeted
Our group's most ambitious project this year!
With tremendous support from the Korean government (aka korean ai-specialized foundation model project), we push the limit of protein structure representation learning.
Very proud of the team behind this work.
What should a structure-aware protein representation be good at?
It should provide geometric signals that structure models can use.
We introduce:
1️⃣ TriProRep: atom-level protein representations
2️⃣ REPSP: a structure-predictive benchmark
bit.ly/3Pm5CHL 👇 (1/9)
Eike Eberhard retweeted
We put the paper online that provides further details (beyond my ICLR keynote) on the role of spontaneous symmetry breaking and Goldstone modes in deep learning. Enjoy! (w/ Nabil Iqbal, Thomas Andy Keller, Takeru Miyato and Yue Song.) arxiv.org/abs/2605.14685
We show that SAIL-trained models leads to significant walltime speedup accross GGA, mGGA and Hybrid functionals, even when extrapolating to 10 times larger molecules
2
68
Joint work with Viktor Kotsev, Timm Güthle and @guennemann
Link: arxiv.org/abs/2604.21657
2
86
Eike Eberhard retweeted
I am hugely excited that our recent work on excited states in neural network QMC has been selected as a spotlight at @icmlconf! 🌟
Huge thanks to Till Grutschus, @FrankNoeBerlin, @guennemann!
arxiv.org/abs/2603.14515v1
Eike Eberhard retweeted
Excited to share that my first PhD paper, which introduces a new diffusion model for relational databases, has been accepted to #NeurIPS2025! We will be presenting it this week in San Diego. ☀️🌴
Joint work with @ludke_david, @SchwinnLeo, and @guennemann.
🧵 1/
Eike Eberhard retweeted
Heading to San Diego for #NeurIPS2025! 🌴☀️
I’ll be presenting 3 recent papers covering generative models for hierarchies, spatiotemporal tissue dynamics, and long-range graph learning.
If you're around, drop by and say hi! 👋
Here’s the schedule 🧵👇
Eike Eberhard retweeted
Our paper “Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow” has been accepted at #NeurIPS2025!
@ale__palmaa @guerrantif @fabian_theis @guennemann
(@TU_Muenchen × @HelmholtzMunich × @MunichCenterML × @MdsiTum )
🧵
Eike Eberhard retweeted
Can we actually tell when LLMs know or don’t know?
For questions with a single answer, that works. But once ambiguity enters - several answers are correct - current methods collapse, confusing model with data uncertainty.
w/ @dfuchsgruber @TomWollschlager @guennemann
[1/4]🧵
Eike Eberhard retweeted
We can sample highly transferable jailbreaks directly from non-autoregressive models (e.g., Diffusion LLMs) without any optimization!
Open-source Diffusion LLMs easily break GPT-5!
In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access.
@TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn
🧵
Eike Eberhard retweeted
Open-source Diffusion LLMs easily break GPT-5!
In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access.
@TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn
🧵
Eike Eberhard retweeted
📣 The Call for Blog Posts for #ICLR2026 is now live:
iclr.cc/Conferences/2026/Cal…
We look forward to your submissions!
Questions? DM me or email: [email protected]






















