New MAGA Framework Improves Cross-Platform GUI Agent Performance

Existing graphical user interface (GUI) agents, often powered by large language models, are typically designed for specific platforms like mobile, web, or desktop. This specialization limits their broader deployment and user experience. The challenge lies in merging these domain-specific experts into a single, versatile agent without corrupting executable actions due to conflicting instructions. Traditional methods like weight merging can lead to inconsistencies, while on-policy distillation often treats all response tokens equally, overlooking the critical role of action tokens in agent-environment interaction. The new MAGA framework addresses this by re-allocating training signals based on the correctness of generated actions. It suppresses invalid distillation signals and prioritizes learning from errors. Additionally, MAGA incorporates a training-only hint mechanism to optimize supervision from specialized teachers without altering the student agent's input. This approach has shown significant improvements, achieving higher mean success rates and outperforming strong baselines across different model scales, demonstrating its effectiveness in creating more robust and adaptable GUI agents.