On-Device Intent Reasoning for Smart Home Agents Via Ontology-Augmented sLLMs
Donghwan Jeong, Honguk Woo · IEEE Access · 2025
Large language model (LLM)-based agents offer powerful natural language understanding for smart home intent interpretation but typically require cloud deployment, raising concerns over privacy, latency, and resource usage. While sLLM can run on-device, they lack the structured contextual knowledge needed to handle ambiguous intents and complex device interactions effectively. To address this limitation, we present CIDER, an ontology-augmented, on-device intent reasoning framework that combines an sLLM with common-sense device knowledge and user-specific context, organized within a structured ontology-based knowledge graph. Specifically, we develop an ontology construction pipeline for smart homes, which leverages cloud-based LLMs to extract and organize common-sense knowledge about smart home devices and integrates user-specific context from a smart home service. The resulting ontology serves as an external knowledge source that augments the sLLM’s intent reasoning process in a retrieval-augmented manner, enabling context-aware inference for autonomous, orchestrated device operation. In evaluations with SmartThings-based testbeds, CIDER achieves a success rate of 82.2% in general scenarios and 76.6% in highly complex home environments, outperforming strong cloud-based baselines by up to 17.7 percentage points, despite relying on a compact on-device sLLM. These results demonstrate that combining the domain-specific efficiency of symbolic knowledge representations with the flexible semantic processing of an sLLM, empowered by an ontology distilled from cloud-based LLMs, provides an effective approach to intent understanding in smart homes.