AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models

Abstract

Smart home assistants must interpret commands ranging from explicit device control to underspecified and preference-dependent requests. Existing Large Language Model (LLM) smart home systems often use heavyweight reasoning pipelines and cloud deployment, limiting efficiency and suitability for resource-constrained environments, raising privacy concerns, and offering limited long-term personalization. We present AdaHome, an adaptive assistant designed for locally deployed Small Language Models (SLMs). AdaHome classifies commands as direct, indirect, or ambiguous and routes them to either a direct planner or a lightweight Chain-of-Draft reasoning planner. It further learns from confirmed and corrected actions using a local semantic and time-weighted preference memory, without prompt-history augmentation or model retraining. Under a unified small model setting, AdaHome achieves 86.7% success on both direct and indirect commands and 88.9% on ambiguous commands, while obtaining the lowest latency and token usage across all categories. In multi-turn evaluation, it achieves 87.5% preference consistency and 100% adaptation success, compared with 52.5% and 30% for a retrieval-augmented baseline.