# UniToMBench Dataset

## Dataset Summary

**UniToMBench** is a unified benchmark designed to evaluate the **Theory of Mind (ToM)** reasoning abilities of large language models (LLMs). It integrates and extends existing ToM benchmarks by providing narrative-based multiple-choice questions (MCQs) that span a wide range of ToM tasks—including false belief reasoning, perspective taking, emotion attribution, scalar implicature, and more.

This dataset supports the research paper [_"UniToMBench: A Unified Benchmark for Advancing Theory of Mind in Large Language Models"_](https://arxiv.org/abs/2506.09450), accepted to NAACL 2025.

---

## Supported Tasks and Annotations

The dataset includes **three subcomponents**, each designed to evaluate different dimensions of Theory of Mind reasoning:

### 1. ToMBench Standard Benchmark
- **File**: `ToMBench_release_v1_0618.xlsx`
- **Tasks**: False belief, first-order/second-order belief, desires, and more
- **Format**: Narrative + MCQ

### 2. Evolving Stories Dataset
- **File**: `evolving_stories_250.xlsx`
- **Format**: Multi-paragraph stories that evolve over time
- **Purpose**: Test long-range mental state tracking

### 3. Multi-Interaction Dataset
- **File**: `multi_interaction_100.xlsx`
- **Format**: Multi-turn dialogue-based scenarios
- **Purpose**: Simulate conversational ToM and dynamic perspective shifts

Each sample contains:
- A story or dialogue
- A question
- Multiple answer options
- The correct answer
- (Optional) metadata such as task type or character labels

---

## Use Cases

- Evaluating general-purpose LLMs on human-like reasoning
- Comparing direct QA vs. multi-step reasoning pipelines
- Fine-tuning or prompting models for ToM-sensitive applications like tutoring agents or social companions

---

## Dataset Structure

Each `.xlsx` file consists of rows formatted as follows:

| Story | Question | Option A | Option B | Option C | Option D | Correct Answer | Task Type | Notes |
|-------|----------|----------|----------|----------|----------|----------------|-----------|-------|

- Formats are standardized across all files.
- Metadata such as `Task Type` and `Notes` help filter or categorize samples.

---

## Citation

If you use this dataset, please cite:

@misc{unito2025,
title={UniToMBench: A Unified Benchmark for Advancing Theory of Mind in Large Language Models},
author={Shamant Sai and Prameshwar Thiyagarajan and Vaishnavi Parimi and Soumil Garg},
year={2025},
eprint={2506.09450},
archivePrefix={arXiv},
primaryClass={cs.CL}
}


Available at: [https://arxiv.org/abs/2506.09450](https://arxiv.org/abs/2506.09450)

---

## License

MIT License