🧠 MeetingToM

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

Paper · GitHub · Project Page

MeetingToM is a multimodal benchmark for evaluating Theory-of-Mind (ToM) reasoning in multi-party meetings. It studies social reasoning at three complementary levels: individual mental states, interpersonal relations, and group-level consensus.

The released benchmark contains 900 source bundles, expanded into 1,200 evaluation records with 1,800 gold answers.

Task	Level	Source bundles	Evaluation records	Questions	Gold answers
STATE	Individual	300	600	Q1: Mental state	600
YOU	Interpersonal	300	300	Q1: Addressee; Q2: Conversational stance	600
CONSENSUS	Group	300	300	Q1: Consensus quality; Q2: Dissenter	600

For STATE, each source bundle contributes two independently evaluated 5-second video clips, resulting in 600 evaluation records. The same mental-state question is asked for every STATE record.

Note on media. MeetingToM is constructed from the AMI Meeting Corpus. AMI-derived video and audio are not redistributed in this repository. Users should obtain authorized access to AMI separately and reconstruct benchmark media locally using the released metadata and scripts.

---

Dataset Viewer and loading

MeetingToM is released with three configurations:

state
you
consensus

Each configuration contains one test split:

state        600 evaluation records
you          300 evaluation records
consensus    300 evaluation records

The configurations can be loaded separately with 🤗 Datasets:

from datasets import load_dataset

state = load_dataset("OliviaWang1101/MeetingToM", "state")
you = load_dataset("OliviaWang1101/MeetingToM", "you")
consensus = load_dataset("OliviaWang1101/MeetingToM", "consensus")

state is the default configuration, so it can also be loaded with:

state = load_dataset("OliviaWang1101/MeetingToM")

---

Tasks

👤 STATE

STATE evaluates whether a model can infer the mental state of a target participant from a short meeting clip.

Each STATE evaluation record contains an independently evaluated 5-second video clip of the target participant together with the corresponding meeting audio.

The model answers the same mental-state question for every clip:

• Q1 — Mental state: infer the target participant's current cognitive or attentional state.

Each source bundle contributes two independent STATE evaluation records, producing 600 STATE records in total.

Example:

{
  "id": "state:ES2002a_state0004_w0",
  "task": "state",
  "bundle_name": "ES2002a_state0004_w0",
  "source_bundle_name": "ES2002a_state0004",
  "session_id": "ES2002a",
  "window_index": 0,
  "answers": {
    "Q1": "COGNITIVE_CONFLICT"
  }
}

👥 YOU

YOU evaluates interpersonal reasoning between meeting participants.

Q1 — Addressee.
The model determines who the current speaker appears to be addressing. Q1 uses the AMI Corner view without audio.

Q2 — Conversational stance.
The model determines the stance expressed by the relevant participant. Q2 uses a 2×2 close-up mosaic with Mix-Headset meeting audio.

Q1 and Q2 refer to the same temporal window.

Example:

{
  "id": "you:ES2002a_you0002",
  "task": "you",
  "bundle_name": "ES2002a_you0002",
  "session_id": "ES2002a",
  "answers": {
    "Q1": "MULTIPLE",
    "Q2": "NEUTRAL"
  }
}

🧩 CONSENSUS

CONSENSUS evaluates group-level social reasoning.

Q1 — Consensus quality.
The model determines the quality of consensus displayed by the group.

Q2 — Dissenter.
The model identifies the participant showing weak buy-in or disagreement when appropriate.

Both questions use the same 2×2 close-up mosaic and temporal window, together with Mix-Headset meeting audio.

Example:

{
  "id": "consensus:ES2002a_consensus0009",
  "task": "consensus",
  "bundle_name": "ES2002a_consensus0009",
  "session_id": "ES2002a",
  "answers": {
    "Q1": "TRUE_CONSENSUS",
    "Q2": "NONE"
  }
}

---

Data format

The three annotation files share several common fields:

Field	Description
id	Unique evaluation-record identifier
task	state, you, or consensus
bundle_name	Released evaluation-record identifier
session_id	Source AMI meeting session
answers	Gold answer dictionary

STATE records additionally contain:

Field	Description
source_bundle_name	Original STATE source-bundle identifier
window_index	Index distinguishing the two STATE records derived from the same source bundle

For STATE, answers contains a single Q1 mental-state label.

For YOU and CONSENSUS, answers contains both Q1 and Q2.

The Hugging Face repository is organized as:

MeetingToM/
├── README.md
├── data/
│   ├── state.jsonl
│   ├── you.jsonl
│   └── consensus.jsonl
├── metadata/
│   ├── reconstruction.jsonl
│   └── reconstruction_summary.json
└── LICENSE-DATA

---

🎬 Reconstruction metadata

metadata/reconstruction.jsonl contains 1,200 reconstruction specifications, aligned one-to-one with the 1,200 released evaluation records.

The metadata provides the information needed to reconstruct benchmark media from an authorized local copy of the AMI Meeting Corpus.

Depending on the task, reconstruction metadata includes information such as:

• AMI session and source timestamps;
• required camera views;
• target participant and view information;
• source-window specifications;
• 2×2 mosaic layout;
• session-specific mappings between close-up views and participant identities.

For STATE, each reconstruction entry corresponds to one independent 5-second video clip.

For YOU-Q2 and CONSENSUS, the close-up mosaic uses the following layout:

Closeup1 | Closeup2
---------+---------
Closeup3 | Closeup4

The reconstruction metadata does not contain gold answers.

Reconstructing benchmark media

The reconstruction code is maintained in the GitHub repository.

After obtaining authorized AMI media, a single STATE evaluation record can be reconstructed with:

python scripts/reconstruct.py \
  --ami_root /path/to/AMI \
  --audio_root /path/to/HeadsetAudio \
  --metadata metadata/reconstruction.jsonl \
  --output_dir reconstructed \
  --id state:ES2002a_state0004_w0

To reconstruct all released evaluation records:

python scripts/reconstruct.py \
  --ami_root /path/to/AMI \
  --audio_root /path/to/HeadsetAudio \
  --metadata metadata/reconstruction.jsonl \
  --output_dir reconstructed \
  --all

Reconstructed AMI-derived media should remain local and should not be redistributed as part of MeetingToM.

---

Evaluation

The official evaluator is maintained in the GitHub repository under:

evaluation/evaluate.py

The benchmark reports the following core metrics:

Task	Metrics
STATE	Accuracy, Macro-F1
YOU-Q1	Accuracy, Macro-F1
YOU-Q2	Accuracy, Macro-F1
CONSENSUS-Q1	Accuracy
CONSENSUS-Q2	Conditional Accuracy
CONSENSUS	Two-step Points Accuracy

STATE metrics are computed over the 600 independent STATE evaluation records.

For CONSENSUS, Q2 is evaluated on records where Q1 is predicted correctly. The point-based score is:

Q1 incorrect                     -> 0 points
Q1 correct, Q2 incorrect         -> 1 point
Q1 correct, Q2 correct           -> 2 points

See the GitHub repository for the complete evaluation protocol and prediction format.

---

Dataset integrity

The public release contains:

Source bundles          900
Evaluation records    1,200
Gold answers          1,800

The 1,200 released evaluation-record IDs align exactly with the 1,200 reconstruction metadata entries.

---

Source media

MeetingToM is based on the AMI Meeting Corpus.

This Hugging Face repository releases:

• benchmark annotations;
• reconstruction metadata;
• reconstruction summary.

It does not release AMI video/audio or reconstructed clips and mosaics.

Users are responsible for obtaining and using AMI data under the applicable AMI access and licensing conditions.

---

Intended use and limitations

MeetingToM is intended as an evaluation benchmark for research on multimodal large language models, Theory-of-Mind reasoning, meeting understanding, and social reasoning.

The benchmark categories are operational annotations for evaluation rather than exhaustive descriptions of human mental states or social behavior. Performance on MeetingToM should not be interpreted as evidence that a model possesses human-like Theory of Mind.

---

Citation

@article{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},
  journal = {arXiv preprint arXiv:2607.19235},
  year    = {2026}
}

---

License

The benchmark annotations and reconstruction metadata in this dataset repository are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.

AMI Meeting Corpus video and audio are not redistributed by MeetingToM and are not covered by this license.

The reconstruction and evaluation code in the GitHub repository is released separately under the MIT License.

---

Authors

Ziyi Wang, Yuhang Wu, Dongxu Piao, Xingyu Liu, Tianhui Zhou, Miao Liu