dailyai.report

Almost none of 22 multimodal models tested via ProactiveBench asked for missing visual information, opting instead to guess. This failure to seek clarification leads to hallucinations in complex tasks. Researchers found a simple reinforcement learning approach fixes this behavior. Practitioners should prioritize these active-learning techniques to improve model reliability in visual reasoning.