ERGO: Efficient High-Resolution Visual Understanding for Vision-Language Models

ERGO (Efficient Reasoning & Guided Observation) is a large vision–language model trained with reinforcement learning on efficiency objectives, focusing on task-relevant regions to enhance accuracy and achieve up to a 3× speedup in inference.

Installation

Python >= 3.10 is required.

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/nota-github/ERGO.git
cd ERGO
uv venv
uv sync
source .venv/bin/activate
uv pip install -e .

Usage

We recommend using vLLM, as its Automatic Prefix Caching can significantly improve inference speed.

This repository provides evaluation scripts for the following benchmarks:

• V* Bench
• HR-Bench
• MME-RealWorld-lite

See data/README.md for how to prepare datasets

Evaluation with vLLM

1. Serving with vLLM

bash ./scripts/run_vllm.sh

2. Run eval.py

export MAX_VISION_TOKEN_NUM=1280
export VLLM_ENDPOINT=http://127.0.0.1:8008/v1
export DATA_ROOT=./data

python ./src/ergo/eval.py \
    --dataset {choose from [vstar, mmerwl, hrbench]} \
    --data_root $DATA_ROOT\
    --api_url $VLLM_ENDPOINT \
    --max_vision_token_num $MAX_VISION_TOKEN_NUM

Inference with Hugging Face

python ./src/ergo/infer.py \
    --input_path {default = "./data/demo/demo.jpg"} \
    --question {default = "Is the orange luggage on the left or right side of the purple umbrella?"} \
    {--save_output} # optional

License

This project is released under Apache 2.0 licence.

Acknowledgements

• We would like to express our sincere appreciation to the following projects:

  • Qwen2.5-VL: The base model we utilized. They are originally licensed under Apache 2.0 License.
  • VLM-R1: The RL codebase we utilized. It is originally licensed under Apache 2.0 License.
  • V*, HR-Bench, MME-RealWorld-lite : The evaluation benchmark dataset we utilized.

• We also deeply appreciate the generous GPU resource support from Gwangju AICA.

Citation

@misc{lee2025ergoefficienthighresolutionvisual,
      title={ERGO: Efficient High-Resolution Visual Understanding for Vision-Language Models}, 
      author={Jewon Lee and Wooksu Shin and Seungmin Yang and Ki-Ung Song and DongUk Lim and Jaeyeon Kim and Tae-Ho Kim and Bo-Kyeong Kim},
      year={2025},
      eprint={2509.21991},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.21991}, 
}