Compress to Focus: Efficient Coordinate Compression for Policy Optimization in Multi-Turn GUI Agents

Yurun Song*, Jiong Yin*, Rongjunchen Zhang, Ian Harris

🔥 News

• [2026-01-12] 🚀 Code and pre-trained models are released!
• [2026-01-XX] 📄 Our paper "Compress to Focus: Efficient Coordinate Compression for Policy Optimization in Multi-Turn GUI Agents" is now available on arXiv.

🚀 Introduction

The official implementation of the paper "Compress to Focus: Efficient Coordinate Compression for Policy Optimization in Multi-Turn GUI Agents".

Abstract: Multi-turn GUI agents enable complex task completion through sequential decision-making, but suffer from severe context inflation as interaction history accumulates. Existing strategies either sacrifice long-term context via truncation or compromise spatial structure through token pruning. In this paper, we propose Coordinate Compression Policy Optimization (CCPO), an efficient policy optimization framework that couples visual compression with policy optimization for multi-turn GUI agents. CCPO introduces Coordinate-Aware Spatial Compression (CASC), which aggregates coordinates from multiple rollouts to capture target-relevant regions and progressively narrow historical attention around key visual areas. From interactions across rollouts, CASC adaptively constructs attention boundaries that concentrate computation on the most informative regions of the scene. We further design a Distance-Based Advantage that provides fine-grained learning signals based on distance rather than binary correctness, improving both grounding accuracy and compression quality. Extensive experiments demonstrate that CCPO achieves SOTA performance across four benchmarks with up to 55% token compression and 3.8$\times$ training speedup.

📈 Method Overview

Overview of the CCPO framework.

✨ Key Features

• Efficient Compression (CASC): Aggregates spatial coordinates to achieve up to 60% token reduction without losing critical context.
• Distance-Based Advantage: Provides fine-grained learning signals based on spatial distance, significantly boosting grounding accuracy.
• Training Acceleration: Delivers 3.5x–4.8x speedup and 16% lower TFLOPS compared to standard RL baselines.
• SOTA Performance: Top-tier results across 4 major benchmarks: Android Control, GUI Odyssey, Mind2Web, and AITW.
• Coupled Optimization: A unified framework that co-optimizes visual focusing and policy decision-making.

🛠️ Installation

Requirements

• Linux
• Python 3.12+
• PyTorch 2.7+
• CUDA 12.8+
• Please refer to requirements.txt for other dependencies.

Setup

# Clone the repository
git clone https://github.com/HiThink-Research/CCPO.git
cd CCPO

# Create a conda environment
conda create -n ccpo python=3.12
conda activate ccpo

# Install dependencies
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

📂 Data Preparation

We evaluate CCPO on four major benchmarks: Android Control, GUI Odyssey, Mind2Web, and AITW, please organize the data as follows:

data/
├── android_control/
├── gui_odyssey/
├── mind2web/
└── aitw/

---

🏃 Usage

1. SFT Training

We first perform Supervised Fine-Tuning (SFT) on Qwen2.5-VL as the warm-up stage.

2. CCPO Training

Then we train the CCPO model with the following command:

cd CCPO
bash scripts/train_CCPO_aitw_7B.sh

2. Evaluation

To evaluate the pre-trained model:

cd ../evaluation
python evaluation_aitw.py \
    --save_path path/to/save/results \
    --model_path path/to/model \
    --his_num 4

---

📊 Model Zoo

We provide pre-trained models (3B and 7B) for reproduction.

Dataset	CCPO-3B	CCPO-7B
AITW	Download	Download

---

📝 Citation

If you find our work useful for your research, please consider citing:

@misc{song2026compressfocusefficientcoordinate,
      title={Compress to Focus: Efficient Coordinate Compression for Policy Optimization in Multi-Turn GUI Agents}, 
      author={Yurun Song and Jiong Yin and Rongjunchen Zhang and Ian G. Harris},
      year={2026},
      eprint={2601.11631},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.11631}, 
}

🙏 Acknowledgement

This project is built upon UI-S1, SimpAgent, and verl-agent. We thank the authors for their great code.