MAGA: A New Method for Training Cross-Environment GUI Agents

A recent study published on arXiv (2607.29320) presents MAGA, a novel technique for training GUI agents that function across mobile, web, and desktop platforms. This research tackles the difficulty of merging domain-specific agents into a cohesive cross-environment policy. Current methods, such as weight merging and on-policy distillation, have their drawbacks: weight merging may distort executable actions due to expert disagreements, while on-policy distillation fails to differentiate, treating all response tokens uniformly. MAGA enhances training signals based on structured actions, minimizing irrelevant or incorrect distillation signals and concentrating on improving flawed actions. Authored by a team of researchers, this paper aims to enhance the deployment and user experience of GUI agents through a unified model across various environments.