---
license: mit
task_categories:
  - other
language:
  - en
  - zh
tags:
  - social-deduction
  - multi-agent
  - vision-language-model
  - game-logs
size_categories:
  - n<1K
---

# QUACK Game Logs

Raw structured event logs (`game.jsonl`) from the QUACK multimodal social deduction benchmark.

## Contents

- **270** games across **9** experimental conditions
- **3** homogeneous (all players same VLM): `gpt5.5`, `claude_opus4.7`, `gemini3.1pro`
- **6** heterogeneous (geese vs duck model pairs)
- **30** seeds per condition (seeds 1–30)

## Layout

```
homogeneous/<model>/<run_dir>/game.jsonl
heterogeneous/geese_<goose>_duck_<duck>/<run_dir>/game.jsonl
manifest.jsonl   # one metadata record per game
```

## Log format

Each `game.jsonl` line is a JSON event (`event_type`, `tick`, `data`, ...).
The `game_started` event includes `initial_state` and `config` for full replay.

See the [QUACK repository](https://github.com/) for evaluation and rendering tools.

## Citation

If you use this dataset, please cite the QUACK paper:
```bibtex
@misc{yuan2026quack,
      title={QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents}, 
      author={Ye Yuan and Rui Song and Weien Li and Zeyu Li and Haochen Liu and Xiangyu Kong and Changjiang Han and Yonghan Yang and Zichen Zhao and Zixuan Dong and Fuyuan Lyu and Bowei He and Haolun Wu and Jikun Kang and Xue Liu},
      year={2026},
      eprint={2605.27068},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.27068}, 
}
```