Twitxy Saved

Wieland Brendel

@wielandbr

Machine Learning Researcher and Social Entrepreneur | Group Lead at ELLIS Institute Tübingen | Co-Founder maddox.ai | Co-Initiator bw-ki.de | @ellis.eu scholar
Tübingen, Deutschland robustml.is.mpg.de Joined August 2015
191 Following 4.3K Followers 507 Posts

Latest posts 5

Pinned post

Wieland Brendel @wielandbr
🚀 @ELLISInst_Tue will lead the AI development for Germany’s new open-source nationwide Adaptive Intelligent System (AIS) learning platform for schools – and we’re #hiring! We are part of a national consortium led by Assecor and @KImachtSchule, and mandated by the FWU.

Join us to push the boundaries of #AI in education and build the core intelligence that will power adaptive learning for millions of teachers and pupils in real classrooms.

We’re developing cutting-edge AI tutoring models, knowledge tracing, and recommendation systems to provide high-quality, personalised learning for all students—regardless of background.

🔹 Several full-time ML roles (2–3 years)
🔹 Top-tier AI research environment
🔹 From prototypes to production-ready solutions for national-scale impact
🔹 Open to on-site, hybrid & remote
👉 Apply now (quick & easy): forms.gle/XmLkwEDD45fY5c…
(or via ais@tue.ellis.eu)

We work closely with our amazing collaborators at Assecor, @KImachtSchule (@stes_io, @SchulzAuguste), Tübingen AI Center (@MatthiasBethge), the Hector Research Institute of Education Sciences and Psychology (Ulrich Trautwein), the Tübingen Center for Digital Education (Andreas Lachner) and the @IWMtue (@ucress).
Let’s shape the future of AI-powered education together. 🔥
#AI #MachineLearning #EdTech #hiring #OpenSource #MLJobs
0 replies 12 reposts 38 likes 3.8K views View on X
Wieland Brendel @wielandbr
Excited to share new work from our group!

Can LLMs truly learn beyond what they were exposed to during pretraining? We introduce LittleLearner, a 5B model pretrained exclusively on a controlled K–5 curriculum.
Shoutout to @fanfei_li7953, @Jana_Zeller, and team!
Fanfei Li @fanfei__li Aug 14
What happens when an LLM never sees material beyond fifth grade?
We trained a 5B LittleLearner model from scratch on LittleCurriculum, a corpus restricted to K–5 material, to make this question testable.🧵
💬 Chat with LittleLearner yourself: littlelearner-ll.github.io
View quoted post on X
0 replies 1 reposts 17 likes 1.5K views View on X
Wieland Brendel @wielandbr
How do we build machines that understand the world like humans?

For over a decade, I've chased that question. It began with adversarial robustness, where imperceptible changes to an image could flip a model's prediction entirely. We went on to show that vision models recognise objects by texture rather than shape. We found that models across machine learning routinely solve tasks through shortcuts. We developed techniques to make them more robust to distribution shifts, and to capture the compositional, object-centric structure of the world. And still, to this day, these models see the world in a profoundly alien way.

Most of that work centred on object recognition. But human vision did not evolve to name objects. It evolved for survival, to let us navigate, grasp, and act in a three-dimensional physical world. Recognition is a byproduct, not the purpose. To understand the world the way we do, a model has to learn how the world actually works: objects, physics, materials, cause and effect. That mismatch, between what we optimise for in the lab and what vision actually evolved to do, has bothered me for years.

Today, together with Georg Martius, Gerard Pons-Moll, Omid Taheri, and Mikel Zhobro, I get to take it on directly. We're launching Ontic Labs: foundation models for physical intelligence — world models that learn how the world works from large-scale video, so any robot can put that understanding to work.

We named it Ontic — that which really, physically exists — because that is exactly the gap we're closing.

The SPRIND - Bundesagentur für Sprunginnovationen selected us as one of ten teams in the €125M Next Frontier AI Challenge to build Europe's frontier AI labs.

That backing lets us build this from Tübingen, one of Europe's strongest AI ecosystems. We wouldn't be here without the Tübingen AI Center and the @ELLISInst_Tue — or without the people behind it like @MatthiasBethge, @bschoelkopf, @GehlerPeter, @pegehler, Alex Diehl and many others.

I can't remember being this excited about a research problem.
Ontic Labs Verified account @onticlabs_io Jul 22
Ontic Labs is Live on X 🚀 and we are hiring!

We are one of 10 teams selected for SPRIND’s €125M Next Frontier AI Challenge to build Europe’s next frontier AI labs.

Our mission is to build foundation models for physical intelligence that learn from human video: both how the physical world behaves and how tasks are carried out inside it.

Our goal is to enable robots to draw on the breadth of human experience and act dexterously, safely, and efficiently across tasks, environments, and embodiments.

Our team:
@omidtaherii, @mikel_zhobro, @wielandbr, @GMartius, @GerardPonsMoll1

Want to build with us? We are hiring exceptional talent in world modeling and foundation models.
More details and opportunities:
onticlabs.io

Supported by @SPRIND

@MPI_IS @uni_tue @ELLISInst_Tue @tubingen @Cyber_Valley

#NextFrontierAI #PhysicalAI #WorldModels #Robotics #Hiring

Photo: Felix Adler für SPRIND
View quoted post on X
Wieland Brendel @wielandbr
Excited to share new work from our group! We introduce MATH-Beyond, a new benchmark for evaluating RL methods. Contains problems that open source base models fail at pass@1024. Success on our benchmark directly measures an RL method's ability to 'expand' a model's reasoning boundaries.
🚀 New Paper & Benchmark!

Introducing MATH-Beyond (MATH-B), a new math reasoning benchmark deliberately constructed for common open-source models (≤8B) to fail at pass@1024!

Paper: arxiv.org/abs/2510.11653
Dataset: huggingface.co/datasets/brend…

🧵1/10
View quoted post on X
0 replies 4 reposts 20 likes 2.6K views View on X
Wieland Brendel @wielandbr
🎯 Come chat with @thwiedemer and @prasannamayil at ICLR'25 today at 10 AM!
They're presenting their spotlight work on CLIP's out-of-domain generalization — poster #570 (Session 1).
Ask them about systematic generalization, reasoning, and world models. 🚀
🎉 New Pre-print! 🎉
Do CLIP models truly generalize to new, out-of-domain (OOD) images, or are they only doing well because they’ve been exposed to these domains in training? Our latest study reveals that CLIP’s ability to “generalize OOD” may be more limited than previously assumed!
Preprint available at: arxiv.org/abs/2410.08258 🧵 1/8
View quoted post on X