LiteGUI: Distilling Compact GUI Agents with Reinforcement Learning

Researchers introduce LiteGUI, a novel training framework that enhances lightweight GUI agents (2B-3B parameters) through reinforcement learning and knowledge distillation, achieving competitive performance with much larger models. The approach addresses key limitations of traditional supervised fine-tuning by incorporating multi-solution learning and dynamic retrieval mechanisms to reduce hallucinations in automated interface interaction tasks.

LiteGUI represents a meaningful advancement in making AI agents practical for on-device deployment, addressing a critical gap in the AI development landscape. Current small-scale models struggle with automated GUI interaction due to training constraints, but this work demonstrates that architectural and methodological innovations can compensate for parameter limitations. The research moves beyond conventional supervised fine-tuning by integrating knowledge distillation with reinforcement learning, allowing 2B-3B parameter models to match performance of substantially larger competitors.

The technical innovation centers on two key contributions: Guided On-policy Distillation leverages oracle reference trajectories to ground agent behavior, while the Multi-solution Dual-level GRPO framework handles the inherent ambiguity in GUI tasks where multiple valid action sequences exist. This dual approach systematically reduces both hallucinations and policy rigidity that plague smaller models. The inclusion of an automated data generation pipeline for multi-solution annotations addresses training data scarcity, a persistent challenge in specialized agent domains.

For the AI industry, this work has tangible implications for edge computing and privacy-conscious applications. Efficient on-device agents enable deployment scenarios where cloud connectivity is unavailable or undesirable, expanding AI agent accessibility across platforms. The demonstrated performance parity between 2B/3B models and larger competitors suggests that future development can prioritize efficiency alongside capability. This shift aligns with market demands for sustainable, deployable AI systems rather than ever-larger foundation models. The structured ablation studies provide a roadmap for practitioners seeking to extract maximum capability from constrained computational budgets, potentially influencing how subsequent agent research prioritizes model scaling versus training methodology sophistication.

• →LiteGUI achieves state-of-the-art performance for lightweight GUI agents (2B-3B parameters) while remaining competitive with much larger models.
• →The SFT-free training paradigm using reinforcement learning and knowledge distillation reduces hallucinations and catastrophic forgetting in small-scale models.
• →Multi-solution dual-level exploration framework specifically addresses the ambiguity inherent in GUI task automation across different action sequences.
• →Automated data generation pipeline with multi-solution annotations enables efficient training of specialized agents without extensive manual annotation.
• →Results demonstrate that structured training methodology can unlock capabilities of constrained models, challenging assumptions about model scaling requirements.