Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs
Simple persona prompting beat more invasive steering methods for making one model answer more consistently across languages.
The paper tests whether an English-prompted Gemma 3 12B Instruct model can be pushed to behave as if queried in German, Spanish, or Bulgarian. It compares persona prompting, Contrastive Activation Addition, and two DPO adapter setups on multilingual factual and culturally rooted benchmark data. Persona prompting is reported as the strongest overall option, while CAA is sharper but fragile and DPO gains are narrower. ArXiv · AI/CL/LG's note
The paper tests whether an English-prompted Gemma 3 12B Instruct model can be pushed to behave as if queried in German, Spanish, or Bulgarian. It compares persona prompting, Contrastive Activation Addition, and two DPO adapter setups on multilingual factual and culturally rooted benchmark data. Persona prompting is reported as the strongest overall option, while CAA is sharper but fragile and DPO gains are narrower. ArXiv · AI/CL/LG's note
score 5