Leveraging Large Language Models for Explainable Activity Recognition in Smart Homes: A Critical Evaluation

LLMs show promise but have significant limitations for activity recognition in smart homes: The paper finds that while Large Language Models can generate natural language explanations more flexibly than traditional rule-based systems, they face challenges when applied to sensor-based activity recognition tasks, suggesting they are not yet a complete replacement for conventional approaches.

Zero-shot LLM approaches require careful consideration: When LLMs are used without training data (zero-shot learning), they can recognize activities based on general knowledge alone, but this comes with trade-offs in accuracy that users should understand before deploying such systems in real smart home environments.

Combining LLMs with existing AI methods may be more practical than replacing current systems: The paper suggests that using LLMs specifically to improve explanation generation—rather than replacing the entire recognition system—could be a more viable path forward, especially when training data is available and high accuracy is important.

Explainability remains crucial for user trust in smart homes: The research emphasizes that as AI systems monitor daily activities in homes, the ability to explain why the system made a particular decision is essential for users to trust and accept the technology, making explanation quality a key responsibility consideration.

Scalability and flexibility of explanations matter for real-world adoption: Traditional sensor-based systems produce rigid, repetitive explanations that users may find unhelpful, while LLMs can generate more natural explanations—highlighting that how AI communicates its reasoning to non-experts significantly impacts responsible deployment.