Jihyoung Jang

VLMs often blunder when faced with ambiguous visual questions, but a new dataset and fine-tuning approach can teach them to strategically seek clarification or list possibilities instead of confidently guessing wrong.