InfiGUI-G1
[AAAI 2026 Oral] Official repository for InfiGUI-G1. We introduce Adaptive Exploration Policy Optimization (AEPO) to overcome semantic alignment bottlenecks in
What is InfiGUI-G1?
[AAAI 2026 Oral] Official repository for InfiGUI-G1. We introduce Adaptive Exploration Policy Optimization (AEPO) to overcome semantic alignment bottlenecks in GUI agents through efficient, guided exploration.
Overview of InfiGUI-G1
What it is
InfiGUI-G1 is a series of models for GUI grounding, accepted as an Oral Presentation at AAAI 2026. The official repository provides the implementation of Adaptive Exploration Policy Optimization (AEPO), a method designed to overcome semantic alignment bottlenecks in GUI agents through efficient, guided exploration.
Key features
- Introduces Adaptive Exploration Policy Optimization (AEPO) to address the exploration bottleneck in GUI grounding.
- Uses a multi-answer generation strategy to explore diverse candidate actions in a single forward pass.
- Employs a theoretically-grounded Adaptive Exploration Reward (AER) function derived from efficiency principles (η=U/C) to balance exploration and exploitation.
- Achieves new state-of-the-art results among open-source models on benchmarks including MMBench-GUI (L2), ScreenSpot-Pro, UI-Vision, UI-I2E-Bench, and ScreenSpot-V2.
- Provides two model sizes: InfiGUI-G1-3B and InfiGUI-G1-7B, publicly available on Hugging Face.
- Includes evaluation and training scripts for reproducibility.
Use cases
- Enhancing GUI agents' ability to ground natural language instructions on computer, phone, and browser interfaces.
- Improving semantic alignment for tasks requiring understanding of abstract visual symbols and implicit instructions.
- Evaluating GUI grounding performance across diverse platforms (mobile, desktop, web) and professional software.
Getting started
Clone the repository and set up the environment:git clone https://github.com/InfiXAI/InfiGUI-G1.gitcd InfiGUI-G1Download models from Hugging Face:mkdir -p ./modelshuggingface-cli download --resume-download InfiX-ai/InfiGUI-G1-3B --local-dir ./models/InfiGUI-G1-3Bhuggingface-cli download --resume-download InfiX-ai/InfiGUI-G1-7B --local-dir ./models/InfiGUI-G1-7BDownload and prepare evaluation datasets, then run evaluation:python eval/eval.py ./models/InfiGUI-G1-3B --benchmark screenspot-pro --tensor-parallel 4
License and pricing
Licensed under Apache-2.0. The project is open source.
Based on the project's public README and repository metadata.