Hidden in Plain Sight: Probing Implicit Reasoning in Multimodal Language Models

This paper analyzes how current multimodal large language models (MLLMs) handle implicit reasoning in real-world, messy environments, revealing that they often fail to detect hidden issues despite possessing relevant skills. Simple inference-time interventions, such as cautious prompting and requesting clarifications, can significantly improve their ability to identify and address implicit problems, highlighting a gap between reasoning ability and behavioral compliance.