ViGaL: Visual Game Learning
Model Overview
We present Visual Game Learning (ViGaL), a novel post-training paradigm where multimodal large language models (MLLMs) develop out-of-domain generalization of multimodal reasoning through playing arcade-like games.
ViGaL-7B demonstrates that training a 7B-parameter MLLM via reinforcement learning on simple arcade-like games like Snake significantly enhances its downstream performance on multimodal math benchmarks like MathVista, and on multi-discipline questions like MMMU, without seeing any worked solutions, equations, or diagrams during RL, suggesting the capture of transferable reasoning skills.
Dataset Usage
Preparing the Training Data
After unzipping the dataset, please check the rotation subfolder.
Converting Image Paths
If you're doing training, you'll need to process the JSON line metadata file in the rotation subfolder. The framework currently only supports absolute image paths, but the JSON line metadata file uses relative paths, so you'll need to add the absolute path prefix.
We provide a simple utility script add_root_prefix.py to convert relative paths to absolute paths. Run this script to update the metadata file before training:
python add_root_prefix.py --input rotation/metadata.jsonl --output rotation/metadata_absolute.jsonl --root /path/to/your/dataset
Running Training
To run the training, please follow the instructions in this README. You can also refer to https://github.com/ModalMinds/MM-EUREKA/tree/qwen for additional information - we're using the same codebase.
Resources
For details of our approach and performance comparison, please see our paper.
For details of training and evaluation, please see our code repo.
| 🚀 Project Page | 📖 Paper | 🔗 GitHub | 🤗 Training Data | 🤗 Model |
Citation
If you find this model useful, please cite our work:
@article{xie2025play,
title = {Play to Generalize: Learning to Reason Through Game Play},
author = {Xie, Yunfei and Ma, Yinsong and Lan, Shiyi and Yuille, Alan and Xiao, Junfei and Wei, Chen},
journal = {arXiv preprint arXiv:2506.08011},
year = {2025},
}