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

AdaHome is a smart home assistant designed around locally deployed small language models. Its intent-aware planning framework routes explicit commands to prompt-based handling and interpretation-heavy requests to lightweight reasoning components, using Chain-of-Draft for efficient decision-making. The paper reports 86.7% accuracy on direct commands, latency reductions of up to 3x, and competitive performance on ambiguous inputs at lower computational cost. For multi-turn personalization, AdaHome reaches 88% preference consistency, compared with 52.5% for a prompt-augmentation baseline, without prompt growth or model retraining.