MeetingToM

Evaluating Multimodal LLMs on
Theory-of-Mind Reasoning in Multi-Party Meetings

Ziyi Wang^1,* Yuhang Wu^1,* Dongxu Piao¹ Xingyu Liu¹ Tianhui Zhou² Miao Liu^1,†

¹ College of AI, Tsinghua University

² Department of Biostatistics and Bioinformatics, Duke University

* Equal contribution † Corresponding author

Explore

Benchmark hierarchy

Subject-level, dyadic-level, and group-level Theory-of-Mind reasoning in multi-party meetings.

MeetingToM task hierarchy showing subject-level, dyadic-level, and group-level tasks — The hierarchical task structure of MeetingToM.

Benchmark overview

MeetingToM evaluates Theory-of-Mind reasoning in multi-party meetings. This section highlights the task structure and a representative example.

Scroll →

MeetingToM task overview showing subject-level, dyadic-level, and group-level tasks

Task structure

Three levels of meeting-grounded social reasoning

The benchmark is organized into subject-level, dyadic-level, and group-level tasks, covering mental states, addressee and attitude, and consensus dynamics.

Example of pseudo-consensus in MeetingToM

Representative example

Beyond surface agreement

Apparent agreement can conceal hesitation or hidden disagreement. The benchmark emphasizes how verbal and non-verbal evidence reveal subtle group dynamics over time.

Citation

The publication venue and final citation will be updated with the paper release.

@misc{wang2026meetingtom,
  title  = {MeetingToM: Evaluating Multimodal LLMs on Theory-of-Mind Reasoning in Multi-Party Meetings},
  author = {Wang, Ziyi and Wu, Yuhang and Piao, Dongxu and Liu, Xingyu and Zhou, Tianhui and Liu, Miao},
  year   = {2026},
  note   = {Preprint}
}

Acknowledgements

MeetingToM is built upon the AMI Meeting Corpus . Original meeting recordings and source materials should be obtained from the official AMI Corpus website.

Additional acknowledgements and funding information will be added with the paper release.