Game2Real-Bench

Game2Real-Bench provides a 1,400-sample multimodal evaluation benchmark and a 23,893-sample game-reasoning training dataset. The two collections are kept separate at the repository root:

• benchmark/ contains the official Toy Game and Real World evaluation data.
• dataset/ contains the training data and its images.

Overview

Collection	Split	Samples	Image files
Benchmark	toy_game	1,000	1,586
Benchmark	real_world	400	779
Training dataset	train	23,893	36,441

The benchmark aligns controlled visual games with real-world tasks across four capability groups: Fine-grained Perception, Spatial Reasoning, Constraint Reasoning, and Sequential Planning.

Repository structure

.
├── benchmark/
│   ├── data/
│   │   ├── toy_game.tsv
│   │   └── real_world.tsv
│   ├── viewer/*.parquet
│   ├── images/
│   │   ├── toy_game/<capability>/
│   │   └── real_world/
│   ├── annotations/toy_game_open.jsonl
│   └── metadata/
└── dataset/
    ├── train.jsonl
    ├── viewer/train.parquet
    ├── images/<task>/
    └── metadata/statistics.json

Loading the data

from datasets import load_dataset

benchmark = load_dataset("liuqa/Game2Real-Bench", "benchmark")
toy_game = benchmark["toy_game"]
real_world = benchmark["real_world"]

training = load_dataset("liuqa/Game2Real-Bench", "dataset", split="train")

The configurations use compact Parquet mirrors so both subsets work in the Hugging Face Dataset Viewer. The benchmark TSV files expose the same concise evaluation fields as the Viewer. The training Parquet mirror exposes the prompt, rules, ordered image paths, options, and canonical answer. Use dataset/train.jsonl for structured game state, solution trajectories, and the deterministic evaluator payload.

For concise browsing, the Viewer renders the first image of each sample. The complete ordered image references remain available in image_path or question_images. Compact 256-pixel previews are embedded only for the Viewer; model inputs should use the separately released original images.

Images are stored separately from the tables. Every image path is relative to the repository root and uses POSIX notation. Preserve the order listed in image_path or public.question_images when constructing multimodal inputs.

To download image assets with the Hub client:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="liuqa/Game2Real-Bench",
    repo_type="dataset",
    allow_patterns=["benchmark/images/**", "dataset/images/**"],
)

Evaluation and answer policy

Evaluation code is maintained in the companion Game2Real-Bench GitHub repository. The Hugging Face repository contains data and annotations only.

The released data uses deterministic canonical answers. For Pipe Rotation, the target minimizes total rotation cost and then selects the lexicographically smallest execution by row, column, and clockwise click count. Rush Hour and Laser answers are replay-validated.

Licensing and attribution

Game2Real-Bench is a mixed-license collection. The repository-level license: other metadata does not replace upstream terms or relicense source material. Review benchmark/metadata/licensing.md, benchmark/metadata/attribution.md, and benchmark/metadata/sources.tsv before using or redistributing the Real World partition. Some sources impose non-commercial or source-specific restrictions.

Citation

Paper: arXiv:2610.18834. If you use Game2Real-Bench or Game2Real-Dataset, please cite:

@article{liu2026game2realbench,
  title   = {Game2Real-Bench: Do Toy Games Reveal and Transfer to Real-World Visual Reasoning?},
  author  = {Liu, Qingan and Feng, Juntong and Zhu, Yue and Cheng, Yujie and Han, Xinzhe and Hao, Shuai and Zhuge, Yunzhi and Diao, Haiwen and Lu, Huchuan},
  journal = {arXiv preprint arXiv:2610.18834},
  year    = {2026}
}