MMToM-QA

MMToM-QA is a multimodal Theory of Mind (ToM) question-answering benchmark and methodology. It challenges large language models (LLMs) to understand human mental states (goals, beliefs, intentions) from diverse multimodal behavioral data. MMToM-QA proposes a new evaluation suite and introduces a model that fuses LLM flexibility with Bayesian inverse planning, advancing the frontier in machine ToM research mainly for academic and AI research communities.