Bennet Wittelsbach @ ICML Testing Workshop

PhD student @TUDarmstadt thinking about probabilistic reasoning in our multimodal world

[TheRealDarthMaul @TheRealDarthMau]

@pfau Can you please ask for a python program (say "A") that takes another program (say "B") as input and:
- if B stops, A should run forever;
- if B doesn't stop, A outputs "Does not stop".

Need it for a friend.

[Chubby♨️ @kimmonismus]

Their newly released pretraining report reveals that Soofi S 30B-A3B is based on NVIDIA’s open Nemotron 3 Nano reference architecture, the same hybrid Mamba + Transformer MoE design with ~3B active parameters and nearly identical architectural choices

What’s actually new isn’t the architecture, but the training:

-~27 trillion training tokens
-German deliberately up-weighted
-End-to-end training on Deutsche Telekom’s Industrial AI Cloud
-Full training recipe, hyperparameters and evaluation methodology released openly

Ngl, excited to see europe kinda trains their own models.

[Dr. Michael Fromm @effi288]

📄 Releasing the Soofi S pretraining tech report: a sovereign, open foundation model for German and English

Today we’re publishing the full pretraining tech report and project page for Soofi S 30B-A3B — a Mixture-of-Experts hybrid Mamba model trained on ~27 trillion tokens with deliberately up-weighted German.

What’s in the report:

🏆 Strongest fully open model in our evaluations on BOTH the English and German aggregates — ahead of Olmo 3 32B and Apertus 70B (full methodology in the report)

📋 Radical transparency: complete per-source data accounting, all hyperparameters, training + eval code, checkpoints — everything under permissive licenses

🇩🇪 Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich — sovereign AI infrastructure on German soil

Soofi S combines frontier-level capability with the highest measured aggregate long-context decode TPS, and unlike full-attention dense baselines maintains high throughput as context grows. The figure plots Capability Index versus measured aggregate decode TPS/GPU at 40K context and batch 32. The Capability Index averages five benchmark groups, i.e., Code, GSM8K, GPQA-Diamond, English aggregate, and German aggregate, after normalizing each group to the best plotted model. Aggregate decode TPS/GPU is measured with a TP=1, one-B200 vLLM latency-subtraction protocol.

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

> 🇩🇪 Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich — sovereign AI infrastructure on German soil

Unexpected German W!

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Paper can be found here! https://t.co/sMQakREDKD

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Really excited about the testing workshop tomorrow! Feel free to stop by my poster "Membership Circuits" to see if we can obtain formal guarantees with learned models, happening at 15:30 🤗

[Feng Liu @AlexFengLiu1]

ICML main conference is coming to an end, but the discussions are not over yet.

Join us at the ICML 2026 Workshop on Hypothesis Testing.
📅 July 11, 2026
📍 Room 318, COEX, Seoul, South Korea

Website: https://t.co/ifMZJU6AAs
Accepted papers: https://t.co/eli8iyRVLh

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Also lots of cool talks, check them out!

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Really excited about the testing workshop tomorrow! Feel free to stop by my poster "Membership Circuits" to see if we can obtain formal guarantees with learned models, happening at 15:30 🤗

[Feng Liu @AlexFengLiu1]

ICML main conference is coming to an end, but the discussions are not over yet.

Join us at the ICML 2026 Workshop on Hypothesis Testing.
📅 July 11, 2026
📍 Room 318, COEX, Seoul, South Korea

Website: https://t.co/ifMZJU6AAs
Accepted papers: https://t.co/eli8iyRVLh

[Alfredo Canziani @alfcnz]

Oh, here is a good one!
«Fast and Expressive Multi-Byte Prediction with Probabilistic Circuits»

[Symposium on Probabilistic Machine Learning @prob_ml]

Excited that our ProbML symposium has now started (and is online on zoom: link is on https://t.co/861JfFwERF)! We have a full day of talks and posters on probabilistic machine learning.

[Reihaneh Zohrabi @ReihaneZb]

Excited to share that HaloProbe is accepted at ICML @icmlconf ! 🎉

Bayesian detection + mitigation of hallucinations in LVLMs (with a Simpson’s paradox twist).

arXiv: https://t.co/B0Z9OikDcA

See you in Korea 🇰🇷
#ICML #ICML2026 #LVLM_Hallucination

[Yujin Jeong @yyujjinii]

Diffusion models fail at multi-object generation — but why? 🤔

In our #ICML2026 paper, we built MOSAIC, a controlled framework to diagnose these failures.

Spoiler: it's not mainly data imbalance. Scene complexity and missing compositions in training matter much more! ✨

(1/n)

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Just arrived in Seoul to join ProbML and ICML. Looking forward to many interesting talks and discussions 😊
(I will also present a poster at the Hypothesis Testing in Machine Learning workshop on saturday, details coming!)

[Martin Mundt @mundt_martin]

Our position paper “Modular Memory is the Key to Continual Learning Agents” (https://t.co/dv77KFe7ll) has been accepted to #ICML2026 @icmlconf as a spotlight! 🎊

Read for a modern perspective on memory, continual learning, and sustainable adaptation at foundation model scale!

[Heng Zhao @Hayouger]

MoE routing shouldn’t be a hard top-k guess.

ProbMoE turns TopK routing into probabilistic inference over expert subsets, giving routers more informative gradients, more exploration, and better expert utilization. Because it’s probabilistic, it naturally extends to Dynamic-k, learning both which experts to use and how many each token needs.

Moreover, ProbMoE is broadly applicable to TopK-routed MoEs, while keeping training efficient through sparse expert execution.

Sparse compute. Smarter routing. 🚀

[Kristian Kersting @kerstingAIML]

So wichtig 👇Wir brauchen einen positiven Blick auf und einen aktiven Zugang zu #KI Modellen in 🇪🇺. Schon lange warnen viele KI-Wissenschaftler vor ungewollten Abhängigkeiten. Glaubt uns endlich. Bitte ! @HolgerHoos

[Stat.ML Papers @StatMLPapers]

Constructing VAE Latent Spaces with Prescribed Topology https://t.co/KEfSZM2Wtc

[Maurice Kraus @ NeurIPS @mkraus_io]

Excited to share KletterMix 🇩🇪🚀

A ~725B-token German pretraining + annealing corpus.

Proud to have co-led this with @HarleRuben, Sebastian Sztwiertnia, Abbas Goher Khan, Mehdi Ali, @effi288, and @kerstingAIML.

Paper: https://t.co/dzg3YQgUyV

[Antonia Wüst @toniwuest]

I'm in Denver at #CVPR2026 this week! 🏔️
Together with @WolfStammer, I will be presenting our recent work, Vision-Language Programs, at Poster Session 3 on Saturday ⏰ 11:45 am - 1:45 pm.
Very excited to share our work and discuss the ideas behind it!

[Bennet Wittelsbach @ ICML Testing Workshop @bennet_wi]

Really enjoyed the AI & Creativity symposium - great talks, great people, and great food ☺️ Thanks for organizing it!

[Antonia Wüst @toniwuest]

Last week, we hosted the Rhine-Main Universities AI & Creativity Symposium in Darmstadt, and the atmosphere was fantastic! 🧠💻

We gathered researchers from many disciplines to explore the big question: What does it truly mean to be creative, for both humans and machines?