## Can LLMs Score Medical Diagnoses and Clinical Reasoning as well as Expert Panels? ##
It appears so... "LLM Jury" has high correlation with the original score of a panel of human experts. Another human panel ("Re-score panel") does not correlate as much.
P: arxiv.org/abs/2604.14892
youtu.be/0qn2z9mD4Tc?...
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
C: github.com/ml-jku/SE-RRM
Mit #OpenClaw organisieren Nutzende ihr Leben. Der KI-Bot braucht dafür allerdings umfassenden Zugriff auf persönliche Daten. Und was hat es mit #Moltbook auf sich, wo sich diese Bots angeblich miteinander unterhalten? @gklambauer.bsky.social, Professor für KI an der JKU Linz, ordnet den Hype ein.
### NEW ENTRY IN THE AI IN DRUG DISCOVERY LEADERBOARD ###
CheMeleon takes the lead as the **best pre-trained model** and gets the Bronze medal (rank 3) overall.
It's an MPNN pre-trained to predict chemical descriptors.
P: arxiv.org/abs/2506.15792
C: github.com/JacksonBurns...
# 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
A NEW ENTRY TO THE TOX21 LEADERBOARD:
GROVER from TENCENT AI LAB obtains rank 5 on the Tox21 leaderboard. Best pre-trained model and clearly outperforms CHEMPROP.
Original implementation: github.com/tencent-aila...
Tox21 leaderboard on Hugging Face: huggingface.co/spaces/ml-jk...
xLSTM for Real-Time DNS Tunnel Detection: arxiv.org/abs/2512.09565
DNS-HyXNet = xLSTM for DNS tunnels.
DNS-HyXNet has 99.99% accuracy, with F1-scores exceeding 99.96%, and per-sample detection latency of just 0.041 ms, confirming its scalability and real-time readiness. wow!
Sepp Hochreiter, Head of LIT AI Lab, comments on the EurIPS conference in this FAZ guest article:
He highlights Europe’s growing visibility in AI research and the role of ELLIS in strengthening the community. 🌍
www.faz.net/pro/digitalw...
Drug discovery benchmarks reveal a surprising result: the original 2017 SELU networks still dominate toxicity prediction (Tox21).
P: arxiv.org/abs/2511.14744
GCN variants (e.g., GyralNet) increasingly replace ReLU with SELU for better convergence and robustness in noisy or shallow graph settings.
P: arxiv.org/abs/2503.19823
SELU made a comeback in RL: PPO-based systems like Pearl report smoother, more stable actor–critic updates when batchnorm is avoided.
P: arxiv.org/abs/2506.01880
The SELU-transformer resurged in NLP/tabular domains -- TabTranSELU even kicks out SwiGLU entirely.
P: doi.org/10.54254/275...
Flowstate continued the trend: SELU activations bring stability in large-scale time-series models.
P: arxiv.org/abs/2508.05287
Time-series foundation models (FIM/FIM-ℓ, Flowstate) all quietly adopt SELU, and TiRex topped the GIFT Eval leaderboard this year.
P: openreview.net/forum?id=NPS...
Tiny SELU MLPs (2–3 layers, width 64) became the standard in Conditional Flow Matching, thanks to smoother derivative behavior than ReLU.
P: arxiv.org/abs/2302.00482
Normalization-Free Transformers rediscover controlled signal propagation — a core idea behind SNNs.
P: arxiv.org/abs/2503.10622
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:
bioinf-jku.github.io/SNNs/
Great news for AI research at JKU!
Upper Austria invests €22.5M to strengthen AI & quantum science, supporting the ELLIS Unit Linz and the LIT AI Lab.
This commitment boosts new projects, talent, collaborations & innovation made in Austria 🇦🇹
www.nachrichten.at/politik/land...
Thanks for more than 1000 visits to the leaderboard in the last 2 days!
YOU THINK YOU HAVE A BETTER MODEL?? GREAT! Training models on the Tox21 Challenge dataset can be done on a laptop-GPU.
Submit your model to the leaderboard! Click SUBMIT here: huggingface.co/spaces/ml-jk...
AI is hitting a wall in drug discovery!
Here is the wall you were all looking for...
huggingface.co/spaces/ml-jk...
Yes, ECFP plus some other descriptors.. should be in the model-cards and the paper ..
You have access to the model predictions via API call.
Just type:
curl -X POST ml-jku-tox21-gin-classifier.hf.space/predict -H "Content-Type: application/json" -d '{"smiles": ["CCO", "c1ccccc1"]}'
in your shell.
Measuring AI Progress in Drug Discovery - A NEW LEADERBOARD IN TOWN
2015-2025: turns out that there's hardly any improvement. AI bubble?
GPT is at 70% for this task, whereas the best methods get close to 85%.
Leaderboard: huggingface.co/spaces/ml-jk...
P: arxiv.org/abs/2511.14744
List of accepted contributions now online: moleculediscovery.github.io/workshop2025/
Roughly 50% of the tickets already gone...
Indeed.. :)
How many errors do you find in the chemical structure above?
Hab mir einen Mahrer Tracker gebaut. Spannend zu sehen, wie schnell Print-Medien auf ihren Online-Portalen auf Dinge wie ein ZIB 2 Interview reagieren und wie sich die Inhalte dabei unterscheiden.
