arxivcs.CLcs.AI2026-07-21
Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs
Although Large Language Models (LLMs) demonstrate remarkable multilingual fluency, their internal knowledge representations remain disproportionately biased toward high-resource languages. This leads to cross-lingual factual inconsistency, where they shift their empirical answer…