Research Sabbatical in Australia: Testing the Safety of "LLäMmlein"
18.03.2026From Brisbane with Insights: How "tame" is our German LLM family really? Our chair returns from a six-week research sabbatical at The University of Queensland with new findings on AI safety and "LLäMmlein".
After a successful six-week research stay at The University of Queensland, Andreas Hotho returns with new international collaborations. Hosted by Gianluca Demartini’s group, the visit focused on bias and safety of large language models (LLMs).
A key starting point was an in-depth exchange on training approaches for models such as LLäMmlein and ModernGBert. What began as a technical discussion quickly led to a surprising insight: LLäMmlein is not as “tame” as its name suggests, showing strong capabilities even in sensitive contexts.
This observation sparked a joint research effort on AI safety, exploring how pretraining data and prompting shape model behavior. While the stay in Brisbane has come to an end, the collaboration continues, with a joint grant proposal already underway.
