Once Correct, Still Wrong: Counterfactual Hallucination in Multilingual Vision-Language Models - Databubble