RAG-HAR: A Training-Free Framework for Human Activity Recognition Using LLMs

Researchers have introduced RAG-HAR, a novel retrieval-augmented generation framework designed for Human Activity Recognition (HAR). Traditional deep learning methods for HAR typically require extensive dataset-specific training, large labeled corpora, and significant computational resources. In contrast, RAG-HAR operates without any model training or fine-tuning. The system computes lightweight statistical descriptors from input data and retrieves semantically similar samples from a vector database. It then leverages Large Language Models (LLMs) to identify activities based on this retrieved contextual evidence. The framework is further enhanced through prompt optimization and an LLM-based activity descriptor that generates context-enriched vector databases, ensuring highly relevant information retrieval. According to the authors, RAG-HAR achieves state-of-the-art performance across six diverse HAR benchmarks. Crucially, it extends beyond recognizing known behaviors, enabling the accurate identification and meaningful labeling of unseen human activities. This approach highlights significant improvements in robustness and practical applicability for applications in healthcare, rehabilitation, fitness tracking, and smart environments, offering a efficient alternative to resource-intensive traditional models.