Title: MiCU: Achieving End-to-End Smart Home Command Comprehension via Large Language Models

Abstract:

Integrating command understanding into smart home platforms can streamline device management and significantly elevate the user experience. Nevertheless, while current systems handle exact instructions—such as "turn on the bedroom light"—with ease, they often falter when faced with vague or context-dependent requests like "make the bedroom cozy." Although Large Language Models (LLMs) demonstrate strong generalization capabilities and typically surpass conventional rule-based approaches in handling such ambiguity, their practical utility is frequently hampered by a lack of specialized data, inadequate task-specific fine-tuning, and substantial computational demands.

To address these challenges, this study introduces MiCU, a specialized LLM designed for superior command interpretation. Our approach begins with an automated pipeline that synthesizes training data by leveraging existing user logs and LLMs. We then implement curriculum learning to embed domain-specific knowledge into the foundational model. To further bolster reasoning capabilities, we apply cold-start training supplemented by reinforcement learning (RL), which is directed by specific thinking protocols tailored to the domain.

Furthermore, we propose a token compression method that summarizes device descriptions into a unique special token. This innovation drastically lowers inference latency and facilitates the creation of \model-fast, a streamlined variant optimized for processing lengthy inputs. Comprehensive evaluations demonstrate that MiCU surpasses baseline models by an average accuracy margin of 20.01% across all device types. Since its deployment in the Xiaomi Home application, the system has garnered roughly 1.7 million daily page views. Operational metrics indicate that MiCU has decreased the user correction rate by 1.57% while boosting human-verified accuracy by 32.05%. The associated code and datasets are publicly accessible at https://github.com/xiaomi-research/iot_spec_llm.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC