The paper “Why are all LLMs Obsessed with Japanese Culture?” turns a vague complaint into a measurement target: when models answer open cultural questions, which countries and regions become their default examples?
The authors introduce Culture-Related Open Questions and report a regional preference pattern that is not limited to the familiar Western or Anglocentric bias frame. They also argue that the effect becomes clearly visible after supervised fine-tuning rather than only during pre-training.
What changed?
Evaluation usually asks whether a model knows facts about many cultures or avoids stereotypes. This paper asks a different question: when the prompt is open, what does the model choose to foreground?
That matters because product behavior is often made of defaults. A model can be broadly competent and still repeatedly route ambiguous cultural questions toward a narrow set of examples.
New Runtime Read
Regional defaults should become part of model evaluation.
For multilingual and global products, the question is not only whether the model can answer in a language. It is whether it distributes attention, examples, recommendations, and assumed context in a way that matches the user population instead of the model’s learned defaults.