Multilingual large language models often give inconsistent answers to culturally ambiguous questions depending on the language used, even when the user persona is fixed. The paper gives the example of a British persona being associated with Shakespeare in English but Cervantes in Spanish, illustrating how the model's cultural grounding shifts with the language of the prompt.

The authors propose a training method called consensus-driven preference optimisation. Rather than treating each language independently, the approach seeks a consensus across languages to guide preference learning, nudging the model toward culturally consistent responses for a given persona.

The work addresses a practical gap in multilingual LLM reliability: users expect a coherent cultural identity regardless of language. The paper does not report specific evaluation results in the abstract, but the proposed method offers a direct way to mitigate this inconsistency during training.