Can MLLMs Read the Room? A Multimodal Benchmark for Assessing Deception in Multi-Party Social Interactions
Researchers introduced Multimodal Interactive Deception Assessment (MIDA), a new task and dataset based on a social deduction game called Werewolf, to test if state-of-the-art Multimodal Large Language Models (MLLMs) can detect deception in complex social settings. The benchmark revealed that current MLLMs struggle to read social cues and model others' beliefs. To fix this, the authors proposed a Social Chain-of-Thought (SoCoT) pipeline and a Dynamic Social Epistemic Memory (DSEM) module, which showed performance improvements.