VideoGlitchBench

A benchmark dataset for game glitch detection in video, containing 5,238 annotated gameplay clips from 120 video games spanning 6 genres. This dataset originates from GamePhysics, a large-scale collection of 26,954 gameplay videos from the GamePhysics subreddit.

Dataset Summary

Statistic	Value
Total videos	5,238
Games covered	120
Videos with glitch	4,717
Videos without glitch (negative)	521
Genres	6 (Action, Simulation, RPG, Strategy, Puzzle, Adventure)

Structure

VideoGlitchBench/
├── <Game Name>/
│   ├── <video_id>.mp4
│   └── ...
└── annotations.json

Videos are organized into subdirectories named after each game. annotations.json contains all labels.

Annotation Format

Each entry in annotations.json corresponds to one video:

{
  "id": 7,
  "game_name": "7 Days to Die",
  "genre": "Action",
  "subgenre": "Survival",
  "video_name": "32akzg.mp4",
  "no_bugs": false,
  "bugs": [
    "A rabbit is moving up and down in the air near the refrigerator door for no reason, indicating a glitch in the physics engine or pathfinding."
  ],
  "time_nodes": [
    [[3, 11]]
  ]
}

Field	Description
id	Unique identifier
game_name	Game title (matches directory name)
genre	Top-level genre
subgenre	Fine-grained genre (e.g., Shooter, Action RPG)
video_name	Filename of the video clip
no_bugs	true if the clip contains no glitch
bugs	List of natural-language glitch descriptions
time_nodes	Temporal segments [[start, end], ...] (in seconds) where each glitch occurs

Genre Distribution

Genre	Videos
Action	3,834
Simulation	974
RPG	313
Strategy	60
Puzzle	47
Adventure	10

Citation

If you use this dataset, please cite:

@article{zheng2026open,
  title={Open-Ended Video Game Glitch Detection with Agentic Reasoning and Temporal Grounding},
  author={Zheng, Muyang and Zhou, Tong and Wu, Geyang and Lin, Zihao and Wang, Haibo and Huang, Lifu},
  journal={arXiv preprint arXiv:2604.07818},
  year={2026}
}