🌐 Project Page | 📃 Paper | 🤗 Playground | 🤗 Evaluation

• [2026.04.07] 🎉 Our CharTide is accepted by ACL 2026 Findings！

V-MAGE is a game-based benchmark designed to evaluate visual-centric capabilities through flexible gameplay and carefully designed levels. Its defining features are as follows:

• Vision Centric Gameplay: Models receive only visual input, requiring pixel-level scene understanding, object tracking, and spatial-temporal reasoning. V-MAGE features continuous-space environments, allowing models to explore the almost infinite state space. Each game is deigned with different difficulty levels that targeting various skill dimensions.

• Extensible Evaluation Framework: V-MAGE extends beyond model evaluation to assess agentic skills that are out-of-scope for current MLLMs. Our game-agent-model three-module evaluation pipeline allows optimizations in both MLLMs and their agent strategies.

• Adaptive ELO-based Ranking: V-MAGE uses a dynamic Elo system to provide a unified and interpretable metric across diverse games and difficulty levels. Unlike raw scores, which vary in scale across tasks, the Elo rating captures relative skill levels by modeling win–loss dynamics between model performances on shared levels.

Performance comparison across different games based on the ELO ranking system. The Random baseline refers to randomly selecting actions from the predefined action space during decision-making phases. Avg. Ratio refers to the average percentage of the model's score compared to the human baseline score.

To evaluate model with V-MAGE, you can use the following steps:

Dependencies can be installed via pip:

cd V-MAGE
conda create -n v-mage python=3.10 -y
conda activate v-mage
pip install -r requirements.txt

If you are using existing api service, you can skip this step.

Otherwise, we recommend using vLLM or SWIFT to deploy the OpenAI interface service for your local model.

Take vLLM and Qwen2.5VL-7B Instruct as an example, you can start the service by running the following command:

# Download the model. 
# Remember to replace <path-to-model> with the path where you want to save the model.
pip install -U huggingface_hub
huggingface-cli download --resume-download Qwen/Qwen2.5-VL-7B-Instruct --local-dir <path-to-model>

# Start the service. You can change the parameters according to your needs.
pip install vllm
vllm serve <path-to-model> --trust-remote-code --max-model-len 15000 --limit-mm-per-prompt image=6 --port 8000 --gpu-memory-utilization 0.90 --tensor-parallel-size 2

You can also use nohup to run the service in the background.

Prepare config file for the model service.

For example, if you are using vLLM, you can simply change the model_path and openai_api_base in the config/model_config/openai_service_config.ini.

[lmm]
model_name = OpenAI
model_path = <path-to-model>
openai_api_key = EMPTY
openai_api_base = http://localhost:8000/v1 # or your own service address

python runner.py \
--llmProviderConfig=./config/model_config/openai_service_config.ini \
--gameEnvConfig=./config/env_config/0steps/env_config_race_reasoning_0steps.json \
--levelConfig=./config/level_config/racegame/level1_no_history.json \
--output_dir=runs/Qwen2_5VL_7B \
--test_rounds=10

python multi_runner.py \
--config_file=./config/multi_runner_config/3steps/Race_3steps.json \
--llmProviderConfig=./config/model_config/openai_service_config.ini \
--output_dir=runs/Qwen2_5VL_7B \
--test_rounds=10

If you don't want to watch the game screen, you can set the environment variable SDL_VIDEODRIVER to dummy before running the script:

export SDL_VIDEODRIVER=dummy

Thanks to the open-source community, we are able to leverage existing game codebases to build our benchmark. Here are the games we used:

@article{zheng2025vmagebenchmark,
      title={V-MAGE: A Game Evaluation Framework for Assessing Visual-Centric Capabilities in Multimodal Large Language Models}, 
      author={Xiangxi Zheng and Linjie Li and Zhengyuan Yang and Ping Yu and Alex Jinpeng Wang and Rui Yan and Yuan Yao and Lijuan Wang},
      journal={arXiv preprint arXiv:2504.06148},
      year={2025},
}