Abstract
Accurate human activity recognition (HAR) is critical for health monitoring and behavior-aware systems. Developing reliable HAR models, however, requires large, high-quality labeled datasets that are challenging to collect in free-living settings. Although self-reports offer a practical solution for acquiring activity annotations, they are prone to recall biases, missing data, and human errors. Context-assisted recall can help participants remember their activities more accurately by providing visualizations of multiple data streams, but triangulating this information remains a burdensome and cognitively demanding task. In this work, we adapt GLOSS, a multi-agent LLM system that can triangulate self-reports and passive sensing data to assist participants in activity recall and annotation by suggesting the most likely activities. Our results show that GLOSS provides reasonable activity suggestions that align with human recall (63-75% agreement) and even effectively identifies and corrects common human annotation errors. These findings demonstrate the potential of LLM-powered, human-in-the-loop approaches to improve the quality and scalability of activity annotation in real-world HAR studies.