InfiX-ai/InfiGUI-G1-7B

InfiGUI-G1-7B: Advanced GUI Grounding with Adaptive Exploration Policy Optimization

InfiGUI-G1-7B is a 7 billion parameter vision-language model developed by InfiX-ai, specifically designed for robust GUI grounding. Fine-tuned from Qwen2.5-VL-7B-Instruct, this model introduces a novel Adaptive Exploration Policy Optimization (AEPO) framework. AEPO is a reinforcement learning method that significantly enhances the model's ability to achieve precise semantic alignment for GUI interactions, addressing the common exploration inefficiencies found in standard RLVR methods.

Key Capabilities

• Enhanced Semantic Alignment: Utilizes AEPO to improve the model's understanding and identification of functionally appropriate UI elements from natural language instructions.
• Efficient Exploration: Integrates a multi-answer generation strategy and a theoretically-grounded Adaptive Exploration Reward (AER) function to dynamically balance exploration and exploitation, leading to more effective learning for complex GUI tasks.
• State-of-the-Art Performance: Establishes new benchmarks among open-source models across challenging GUI grounding datasets.

Good For

• Developing GUI Agents: Ideal for building intelligent agents that can accurately interpret and interact with graphical user interfaces based on natural language commands.
• Automated UI Testing: Can be used to automate testing processes by grounding instructions to specific UI elements.
• Accessibility Tools: Potentially applicable in creating advanced accessibility solutions that translate user intent into precise GUI actions.
• Research in GUI Grounding: Serves as a strong baseline and research tool for advancing the field of vision-language models in human-computer interaction.