MobileRLHF: Towards Privacy-Preserving and Self-Improving LLM-Based Local Health Agents on Smartphones via Federated RLHF

Abstract

Mobile and wearable sensing enables health agents to provide personalized insights from users' health data, but existing cloud-hosted LLM backends raise privacy concerns. We present MobileRLHF, a privacy-preserving and self-improving framework that enables human-feedback-driven preference adaptation directly on commodity smartphones through federated RLHF. MobileRLHF addresses two key challenges: resource constraints of mobile devices, and the sparsity of personal feedback. To the best of our knowledge, MobileRLHF is the first practical system to execute human-feedback-driven LLM preference optimization directly on commodity smartphones. We implement MobileRLHF with a real-world federated deployment, and conduct preliminary experiments to evaluate local and federated adaptation behavior and the quality of agent outputs. Our code is open-sourced at: https://github.com/Edge-Intelligence-Lab/MobileRLHF.