LiteGUI: Reinforcement Learning Distills Compact GUI Agents
A new training paradigm for on-device GUI agents, LiteGUI, uses reinforcement learning and knowledge distillation to improve small-scale vision-language models without supervised fine-tuning. The method, Guided On-policy Distillation, integrates oracle trajectories and dynamic retrieval to reduce hallucinations and cognitive misalignment in multi-solution GUI tasks. The paper is published on arXiv (2605.07505).
Key facts
- LiteGUI is a training paradigm for on-device GUI agents.
- It uses reinforcement learning and knowledge distillation.
- It avoids supervised fine-tuning (SFT).
- Guided On-policy Distillation is the core method.
- It incorporates oracle reference trajectories.
- It uses a dynamic retrieval mechanism.
- It reduces hallucinations and cognitive misalignment.
- The paper is on arXiv with ID 2605.07505.
Entities
Institutions
- arXiv