The Bethge Lab is a machine learning research group at the University of Tübingen and the Tübingen AI Center, led by Matthias Bethge, and is an academic, non-profit research team. The lab works in an open manner and collaborates closely with the non-profit organization Open Athena, which is dedicated to bringing frontier AI capabilities into academia.
Main research directions The lab's core theme is "continual self-improvement," that is, studying how intelligent agents get better through use. Specifically, it focuses on how AI systems, after being deployed into the hands of real users, can continue learning through mechanisms such as in-context learning, memory, merging, and fine-tuning on use, rather than being limited to laboratory environments.
The lab lists three key research directions:
- Recursive Self-Improvement: Can AI systems build subsequent AI systems?
- Human Learning in the Age of AI: How can people best use continually evolving AI agents?
- Foundation Models for Neuroscience: What can models trained on every neuron we record teach us about the brain?
Representative research outputs The lab has publicly released multiple studies, including:
- PostTrainBench: Can LLM agents automate LLM post-training? It explores whether AI agents can fully autonomously post-train AI models.
- AlgoTune: Can language models accelerate general numerical programs? It explores whether AI agents can invent genuinely faster algorithms.
- ONEBench: Sample-level benchmarking for open-ended capabilities, allowing evaluation to grow together with the models being tested.
- QVal: A low-cost dense supervision signal for evaluating long-horizon LLM agents, teaching agents to identify which steps are truly helpful, thereby learning from long-horizon real-world tasks.
- DeepGaze3.5-VL: Modeling gaze paths through autoregressive token prediction, using today's foundation models to predict where peoplelook.
- openretina project: Collaborative retina modeling across datasets and species, jointly building openly shared retina models with multiple labs and species.
Collaboration and talent development The lab collaborates closely with Open Athena. Doctoral students can join through IMPRS-IS and ELLIS. The lab openly publishes its benchmarks, code, and data. It is currently recruiting doctoral students, postdocs, and research engineers.
Applicable scenarios and limitations The lab is mainly oriented toward academic research and frontier exploration, and is suitable for researchers, students, and collaborators interested in continual learning of AI agents, recursive self-improvement, AI evaluation benchmarks, and foundation models for neuroscience. Its outputs are released in the form of open benchmarks, code, and data, and can be used by the academic community. It should be noted that the lab is an academic non-profit research group and does not provide commercial products or services; information about specific fees, prices, platform support, and similar details is not currently provided in the materials.
