🧩 VRQA-Bench
A multiple-choice visual reasoning benchmark built on 🌀 maze navigation and 📦 Sokoban puzzles. Each sample provides an initial-state image, an optimal-solution video, and a question about the planned trajectory.
🏷️ Categories
| Category | Task | Description |
|---|---|---|
C1_turn_count |
🌀 maze | Total number of direction changes along the path |
C2_turn_direction |
🌀 maze | Number of LEFT / RIGHT turns |
C4_sokoban_push |
📦 sokoban | Direction of a specific push |
C5_direction_count |
🌀 maze | Number of steps in a given direction |
C6_push_dir_count |
📦 sokoban | Number of pushes in a given direction |
📝 Annotation Fields
question, options (A–D), answer (correct key), correct_value, task_type, input_image / video_path (relative to each split dir), _solution (ground-truth solution trace).
📌 Citation
@misc{zhou2026worldmodelsmeetlanguage,
title={World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning},
author={Yucheng Zhou and Wei Tao and Yiwen Guo and Jianbing Shen},
year={2026},
eprint={2606.03603},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2606.03603},
}