ICML 2026
PostTrainBench: Can LLM Agents Automate LLM Post-Training?
Can AI agents post-train AI models entirely on their own?
We study continual self-improvement — agents that get better with use.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
Bethge Lab
We study continual self-improvement — agents that get better with use.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
We study continual self-improvement — agents that get better with use.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
We study continual self-improvement — agents that get better with use.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
We study how AI
systems keep learning
after deployment.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
The Bethge Lab is an academic, non-profit research group at the Tübingen AI Center, led by Matthias Bethge.
We work in the open, closely with Open Athena, the non-profit bringing frontier AI capabilities to academia.
Research
More on our research →Selected publications
All publications →NeurIPS 2025
AlgoTune: Can Language Models Speed Up General-Purpose Numerical Programs?
Can AI agents invent genuinely faster algorithms?
ACL 2025
ONEBench to Test Them All: Sample-Level Benchmarking Over Open-Ended Capabilities
Evaluation that keeps growing with the models it measures.
NeurIPS 2026
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
Teaching agents which steps actually helped, so they can learn from long, real-world tasks.
ECCV 2026
DeepGaze3.5-VL: Modeling Scanpaths via Autoregressive Token Prediction
Predicting where people look with today's foundation models.
eLife 2025
The openretina Project: Collaborative Retina Modelling Across Datasets and Species
Building an open, shared model of the retina together, across labs and species.
The lab
We collaborate closely with Open Athena. PhD students join us through IMPRS-IS and ELLIS. We publish our benchmarks, code and data openly.
People
Current members and alumni →Join & contact
We are recruiting PhD students, postdocs and research engineers, and are happy to hear from you.