PhoneBuddy: Training Open Models for Agentic Phone Use (phonebuddyai.github.io)

PhoneBuddy has introduced a novel approach for training open phone-use agents by integrating real-app reinforcement learning (RL) with scalable mock-app environments. This combination allows the system to leverage the realism of executing tasks in actual applications while also benefiting from the flexibility and verifiability of mock-app training. The key finding demonstrates that this mixed-training approach significantly enhances the capabilities of the PhoneBuddy-4B-Real+Mock model, particularly in single-app tasks and AndroidWorld environments, achieving higher success rates compared to existing models like GPT-5.4 and Gemini 3.1 Pro. This development is significant for the AI/ML community as it paves the way for more effective and reliable mobile AI agents. The scalability of training pipelines translates to better generalization in varied contexts, bolstering safety and privacy protocols through robust benchmarking and verification of agent behaviors. However, a noted limitation remains in cross-app interactions, suggesting that while single-app performance is strong, the ability to seamlessly transition between multiple applications needs further refinement. PhoneBuddy sets a promising foundation for future advancements in mobile AI interactions.