MAPLE: A Multimodal Mobile Agent with Playbook-Learned Experience
Jun Wang, Lucas Spangher, Eugene Ie and colleagues at Google and Washington University in St. Louis present MAPLE, a training-free multi-agent framework for mobile GUI agents that writes reusable rules into a shared playbook at every step instead of reflecting only after an episode ends.
Ask this paper
Architecture. A bounded working memory tracks state inside an episode. A persistent playbook stores declarative rules about UI behavior distilled in hindsight, deduplicated, and retrieved with the task intent masked so that rules transfer across tasks.
Results. With Gemini-3-Flash, MAPLE reaches 75.9% success on AndroidWorld against 71.6% for the M3A baseline on the same model, and 60.25% on MobileWorld, including 68.18% on interactive tasks that require asking the user.
Scaling. Across six vision-language models, the gains are largest for small and open-weight models, where the scaffolding prevents errors from accumulating under ambiguous instructions.
Limits. The authors report a scaffolding paradox in which frontier models gain less from the structure, and the method still trails unrestricted leaderboard systems that combine several large models.
Abstract
Autonomous agents powered by vision-language models (VLMs) show great potential in navigating mobile graphical user interfaces (GUIs) to complete complex tasks. However, existing paradigms face fundamental scalability bottlenecks: fine-tuning requires massive high-fidelity datasets or prohibitive training time and compute, while gradient-free approaches rely on custom static heuristics or costly trial-and-error trajectory exploration where feedback is deferred until the end of execution. To address these limitations, we introduce MAPLE (Multimodal Agent with Playbook-Learned Experience), a novel gradient-free multi-agent framework that enables dynamic, step-wise experience sharing and execution-time prompt evolution. Unlike traditional action-reflection architectures that depend on costly, post-hoc self-reflection, MAPLE enables agents to cross-reference and distribute runtime insights in real time at each operational step. This dynamic exchange eliminates blind trial-and-error looping and steers agents along direct, grounded interaction paths. Extensive evaluations on the AndroidWorld and MobileWorld benchmarks demonstrate that MAPLE achieves state-of-the-art performance on complex, long-horizon mobile tasks, while proving scalable and adaptable across novel environments.