Researchers have introduced MIDA, a benchmark designed to test whether multimodal large language models (MLLMs) can assess deception in multi-party social interactions. The dataset provides synchronized video and text, with every statement labeled as truthful or deceptive, allowing for rigorous evaluation of models' social reasoning.

Benchmarking 12 state-of-the-art open- and closed-source MLLMs, the authors found a substantial performance gap: even powerful models like GPT-4o struggle to reliably identify deceptive statements. The analysis suggests current models have difficulty grounding language in multimodal social cues and modeling what others know, believe, or intend.

To address these limitations, the paper proposes a Social Chain-of-Thought (SoCoT) reasoning pipeline and a Dynamic Social Epistemic Memory (DSEM) module. Experiments show that structured epistemic memory consistently improves deception assessment, while multimodal social-cue reasoning remains a challenging direction for future work. The paper has been accepted to EMNLP 2026.