Most strikingly, she emphasises that just a few examples can cause far-reaching behavioural shifts in LLMs, potentially affecting current models as well. For practitioners, the takeaway is clear: careful training data curation and thorough testing after fine-tuning are essential.
(3/馃У )
In a new briefing by the Science Media Center Germany, Prof. Dr. Iryna Gurevych (Ubiquitous Knowledge Processing Lab, Technische Universit盲t Darmstadt) notes that the study鈥檚 methodology is well aligned with its claims: It extends earlier work by the same lab showing that fine-tuning can lead to broader misalignment.
(2/馃У )
Most strikingly, she emphasises that just a few examples can cause far-reaching behavioural shifts in LLMs, potentially affecting current models as well. For practitioners, the takeaway is clear: careful training data curation and thorough testing after fine-tuning are essential.
(3/馃У )
The briefing also features perspectives from:
馃懁 Prof. Dr. Hinrich Sch眉tze, Ludwig-Maximilians-Universit盲t M眉nchen (LMU)
馃懁 Prof. Dr. Dorothea Kolossa, Technische Universit盲t Berlin
馃懁 Dr. Paul R枚ttger, Oxford Internet Institute
馃懁 Dr. Jonas Geiping, Max Planck Institute for Intelligent Systems
馃搫 Read the full German briefing here:
https://sciencemediacenter.de/angebote/sprachmodelle-entwickeln-unerwuenschte-verhaltensweisen-26006
馃Ь Nature paper:
https://www.nature.com/articles/s41586-025-09937-5
(4/4)
