InfiX-ai/InfiGUI-G1-3B
InfiGUI-G1-3B is a 3 billion parameter vision-language model developed by InfiX-ai, fine-tuned from Qwen2.5-VL-3B-Instruct. It utilizes Adaptive Exploration Policy Optimization (AEPO) to enhance semantic alignment for GUI grounding tasks. This model excels at accurately identifying and interacting with UI elements based on natural language instructions, establishing new state-of-the-art results among open-source models on various GUI grounding benchmarks.
Loading preview...
InfiGUI-G1-3B: Advanced GUI Grounding with AEPO
InfiGUI-G1-3B is a specialized vision-language model from InfiX-ai, built upon the Qwen2.5-VL-3B-Instruct architecture. Its core innovation lies in the Adaptive Exploration Policy Optimization (AEPO) framework, a novel reinforcement learning method designed to significantly improve the model's semantic alignment for complex Graphical User Interface (GUI) grounding tasks.
Key Capabilities & Innovations
- Enhanced GUI Grounding: Addresses the challenge of robustly grounding natural language instructions to GUI elements, focusing on both precise spatial and correct semantic alignment.
- Adaptive Exploration Policy Optimization (AEPO): Overcomes exploration bottlenecks common in standard Reinforcement Learning with Verifiable Rewards (RLVR) methods. AEPO integrates a multi-answer generation strategy for diverse action exploration, guided by a theoretically-grounded Adaptive Exploration Reward (AER) function.
- State-of-the-Art Performance: Achieves leading results among open-source models across multiple challenging GUI grounding benchmarks, including MMBench-GUI (73.4), ScreenSpot-v2 (91.1), UI-Vision (22.0), I2E-Bench (72.6), and ScreenSpot-Pro (45.2).
Ideal Use Cases
- Automated GUI Interaction: Developing agents that can understand and interact with user interfaces based on natural language commands.
- UI Element Identification: Precisely locating and identifying functional UI elements from visual inputs.
- Improving Semantic Alignment: Applications requiring high accuracy in associating natural language instructions with the correct semantic meaning of GUI components.
- Research in RL for GUI: A strong baseline and framework for further research into reinforcement learning techniques for GUI agents.