Pinned Tweet
Yesterday, I successfully defended my PhD dissertation on 'Multimodal Contrastive Learning for Drug Discovery' 🎓. I'm proud of our work, which demonstrates how to exploit multiple modalities in drug discovery using contrastive learning and modern Hopfield networks.
Philipp Seidl retweeted
New paper 🧵 KronLSTM: a new member of the xLSTM family, built on the Kronecker Delta Rule.
✏️ less interference than delta-rule updates
📏 provably bounded memory, no normalizer state
📈 competitive with, often ahead of, (Gated) DeltaNet, especially at long context
Philipp Seidl retweeted
KronLSTM is out!! | Candidate to replace TRANSFORMER in LLMs
KronLSTM is built on the Kronecker Delta Rule: instead of correcting the vector stored at a key, it corrects a single scalar, how strongly this key is bound to this value.
P: arxiv.org/abs/2610.26834
Philipp Seidl retweeted
GPT-6 Astra⭐️outperformed 2 specialist models (LocalRetro and RetroKNN) in single-step retrosynthesis prediction. Unbelievable result just a month ago! It is the very first general-purpose model outperforming these stong specialists in our benchmark. Kudos to @gdb @joyjiao12 👏
In the #DDDBenchmark, the model ranks 3rd🥉overall, behind only 2 other specialist models including our model and MHNReact by @gklambauer @phseidl , while significantly outperforming all other frontier models (see in thread⬇️).
Btw, this synthesis visualization itself was also generated by GPT-6 Astra in just 15 minutes. We will see an explosion of 3D simulations of reactions thanks to models like Astra! 🧵1/2
#insilicoSOTAFM
Standard exact-match fails to capture the one-to-many complexity of retrosynthesis. A new preprint (arXiv:2608.18940) introduces ChemCensor to evaluate Top-K chemical plausibility, showing that generative LLMs and template-based models (MHNreact) explore complementary spaces. 🧪
1
4
264
Philipp Seidl retweeted
LLMs can only think out loud: every thought is a token, every token a forward pass.
We trained LLMs to reason in a latent working memory instead. No reasoning trace, TTFT matches direct answering.
See it for yourself in my new interactive blog post:
👉 aichberger.github.io/blog/re…
Philipp Seidl retweeted
New paper "On Subquadratic Architectures: From Applications to Principles" 🙌
On commonsense & reasoning benchmarks, xLSTM, Mamba-2 & Gated DeltaNet perform nearly indistinguishably.
Therefore, we analyse them where structure genuinely matters: on code & time series.👇
Philipp Seidl retweeted
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!
Philipp Seidl retweeted
AI for biomedical imaging breaks when you change the lab, device, or batch.
New paper fixes this using the control samples already in your experiment. The forgotten controls (and meta-learning) were the answer all along.
P: arxiv.org/abs/2604.20824
Philipp Seidl retweeted
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
Philipp Seidl retweeted
🎉 Excited to share that our paper, AP-OOD, has been accepted at @iclr_conf!
We introduce Attention Pooling for Out-of-Distribution Detection:
AP-OOD detects OOD inputs in the latent token space of Transformer models, drastically improving results on natural language tasks.📈
Philipp Seidl retweeted
🏆 MolecularIQ is live — and open to the community
👉 Check how current LLMs perform on real molecular structure reasoning
👉 Submit your own chemistry LLM and get evaluated under a standardized protocol
🔗 Leaderboard & submissions:
huggingface.co/spaces/ml-jku…
Philipp Seidl retweeted
ELLIS ML4Molecules Workshop running at the Ellis Unconference at #EurIPS, Copenhagen.
Philipp Seidl retweeted
Interested in how we can use ideas from Flash Attention for more efficient linear RNN kernels?
I am heading to NeurIPS in San Diego to present our work on Tiled Flash Linear Attention: More Efficient Linear RNN and xLSTM Kernels.
Philipp Seidl retweeted
🚀 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
Philipp Seidl retweeted
A devastating blow for GNNs/MPNNs in chemistry...
P: arxiv.org/abs/2508.06199
"All models, except CLAMP, are either worse than or practically equivalent to the baseline ECFP fingerprint on molecular property prediction tasks." from Benchmarking Pretrained Molecular Embedding Models arxiv.org/pdf/2508.06199 - honored =)
"All models, except CLAMP, are either worse than or practically equivalent to the baseline ECFP fingerprint on molecular property prediction tasks." from Benchmarking Pretrained Molecular Embedding Models arxiv.org/pdf/2508.06199 - honored =)
Specifically trained reasoning LLM on retrosynthesis data (trained on USPTO-full wo 50k-test, though)
2
165






















