Mark RadarMARK RADAR
中 EN
Event File AI Large Language Models (LLMs)

Study Links AI Models’ Japan Bias to Supervised Fine-Tuning

1 reports · First detected 2026-07-23 · Last active 2026-07-23

Large language models increasingly shape search, education and content production, making cultural skew a practical concern for global users. Earlier research largely emphasized Western or Anglocentric bias; the new study instead used open-ended cultural questions without named locations, forcing models to choose a country or region. The approach is designed to expose which cultures models treat as salient when several answers could be valid, a key test of whether AI systems represent diverse societies evenly.

Researchers at the HiTZ Center-Ixa of the University of the Basque Country EHU and Cardiff University posted the paper to arXiv on April 23, 2026. They built CROQ, comprising 31,680 questions across 24 languages, 11 themes and 66 subtopics, then tested eight frontier models. Japan was the most frequently cited foreign country in six models. Comparisons of base, supervised fine-tuned and instruction-aligned OLMo checkpoints found the biggest concentration shift during SFT, pointing to alignment data rather than raw pretraining corpora as the main source.

All Coverage

1 original reports
TECHNEWS.TW 2026-07-23
研究揭密：主流模型為什麼獨鍾日本文化

The Backstory

The history behind this event
Study Finds Large Language Models Culturally Biased Toward Japan and the US2026-05-05 · 1 reports · similarity 0.90

Large language models have become gateways to information across languages, but their training data and fine-tuning processes may narrow the range of cultural perspectives they offer. Researchers at the University of the Basque Country in Spain and Cardiff University in the UK therefore created a set of open-ended cultural questions to examine whether models default to a small number of dominant cultures when answering questions about food, dance and other topics that do not specify a country, potentially limiting the diversity of content available to users worldwide.

The team released its paper on April 23, 2026. Its CROQ dataset contains 31,680 questions across 24 languages and was used to test eight models, including GPT-4o-mini and Gemini 2.5 Flash. After excluding countries associated with the language of each question, six models cited Japan most often, while the US also ranked near the top. Additional experiments showed that the bias emerged mainly after supervised fine-tuning rather than during pretraining.

Mark Radar｜MARK RADAR
Past Issues・by MarkReadFintech
All times are in Taipei time (GMT+8)