SignalTech/AI
GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning
GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning
Context
What happened: Researchers introduced GroundGUI, built on the ScreenRef GUI grounding dataset and a Distillation-to-Reinforcement Learning (D2RL) method, training GroundGUI-8B from Qwen3-VL-8B-Instruct via sequential SFT and D2RL.
Key numbers: ScreenRef contains 586K screenshots and over 633K referring expression instances.
Why it matters: The authors report state-of-the-art results among 8B-scale agents, with performance comparable or superior to models trained on closed-source data.
What changed
Researchers introduced GroundGUI, built on the ScreenRef GUI grounding dataset and a Distillation-to-Reinforcement Learning (D2RL) method, training GroundGUI-8B from Qwen3-VL-8B-Instruct via sequential SFT and D2RL.
Why it matters
The authors report state-of-the-art results among 8B-scale agents, with performance comparable or superior to models trained on closed-source data.