BridgeAI-Lab/omnitom-benchmark

OmniToM

OmniToM is a benchmark for evaluating Theory of Mind in language models through explicit belief-structure modeling. Many benchmarks ask a model to answer a question about a social story. OmniToM provides annotated belief structures for the stories, so researchers can compare a model's extracted beliefs with what the benchmark records about each person.

Each example pairs a short English story with actor-centered belief statements and seven labels for each belief. The benchmark supports two linked tasks:

1. Belief extraction: Given a story, identify the actors, their belief statements, and the order of each belief.
2. Belief labeling: Given a story and its belief statements, assign the seven schema labels to each belief.

The released data contains the annotated stories and beliefs. The paper explains the benchmark design and evaluation protocol.

Dataset Summary

Measure	Count
Stories	895
Story categories	7
Belief statements	22,343
Labels per belief	7
Total belief labels	156,401

The stories come from seven ToMBench categories:

Story category	Stories
Ambiguous Story Task	98
False Belief Task	97
Faux-pas Recognition Test	142
Hinting Task Test	100
Persuasion Story Task	97
Scalar Implicature Test	154
Strange Story Task	207
Total	895

Data Fields

The release contains one JSONL file, benchmark_story_belief_labels.jsonl. Each line is one story. Its beliefs list contains the belief statements and their labels. The reserved actor world is used for facts stated by the narrator rather than beliefs held by a character.

Field	Description
story_id	Unique story ID.
story_category	ToMBench story category.
story	Story text.
beliefs	List of annotated belief statements.
beliefs[].actor	Belief holder. world denotes a narrator or world fact.
beliefs[].belief	Belief statement.
beliefs[].labels.order	Belief depth: 0, 1, 2, or 3.
beliefs[].labels.truth_status	True, False, or Unknown.
beliefs[].labels.knowledge_access	Private, Shared, or Public.
beliefs[].labels.representation	Explicit or Implicit.
beliefs[].labels.content_type	Location, Contents/Physical State, Identity/Relation, Epistemic, Desire/Intention, Emotion, Trait/Value, or Action/Event.
beliefs[].labels.mental_source	Narration, Perception, Memory, Testimony, Inference, Imagination, or Unknown.
beliefs[].labels.context	Neutral, Temporal, Deceptive, or Counterfactual.

Usage

from datasets import load_dataset

stories = load_dataset("BridgeAI-Lab/omnitom-benchmark", split="train")
print(stories[0])

Hugging Face calls the single released benchmark split train for loading purposes. It is the benchmark split, not a recommended training set. The code repository contains the evaluation scripts.

Source and Limitations

The stories are drawn from ToMBench. The belief statements and labels were produced through an LLM-assisted process with human validation and adjudication, as described in the OmniToM paper. This release contains the seven retained story categories, rather than the full ToMBench collection.

OmniToM covers English, text-only, short stories. Its labels may reflect judgment calls in ambiguous social situations. The dataset is best used to study belief representations in these stories, not as a general measure of real-world social reasoning.

Citation and Resources

@misc{omnitom2026,
  title={OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling},
  author={Bawatneh, Adam and Sapkota, Sagar and Bedi, Amrit Singh and Karmaker, Santu and Shah, Mubarak},
  year={2026},
  eprint={2605.26322},
  archivePrefix={arXiv}
}

Project page · Code · Croissant metadata with Responsible AI fields