Bethge Lab

Bethge Lab is a university research group based in Tübingen, Germany, working at the intersection of machine learning, computational neuroscience and AI safety-adjacent questions of self-improvement. Its stated research theme is continual self-improvement: how agents can improve in-context, through memory, model merging or fine-tuning on use, so that they keep getting better after deployment in real users' hands rather than only in the lab.

The group's work is organised into three broad threads: recursive self-improvement (can AI systems help build the AI systems that come after them), human learning in the age of AI (how people can best work with continuously evolving agents), and foundation models for neuroscience (what large models trained on recorded neural activity can teach us about the brain). Representative outputs include PostTrainBench, which probes whether LLM agents can automate LLM post-training; AlgoTune, which tests whether language models can invent faster numerical algorithms; ONEBench for sample-level benchmarking across open-ended capabilities; QVal for cheaply evaluating dense supervision signals for long-horizon agents; DeepGaze3.5-VL for scanpath prediction; and the openretina project for collaborative retina modelling across labs and species.

The lab sits within the University of Tübingen, the Tübingen AI Center and the ELLIS Institute Tübingen, and collaborates closely with the non-profit Open Athena. It publishes benchmarks, code and data openly, recruits PhD students, postdocs and research engineers, and positions itself as an academic, open-source effort rather than a vendor.