TRACE-TS: New Framework for Signal-Grounded Reasoning in Wearable Sensor Data

A recent research article presents TRACE-TS (Traceable Reasoning with Attribution-Grounded Evidence), a framework aimed at enhancing the interpretability of human activity recognition using data from wearable sensors. This paper, found on arXiv (2608.00200), tackles a significant drawback in existing language model (LM)-driven methods that produce natural-language explanations for sensor data, which often lack a connection to the actual signals, leading to articulate yet unverifiable narratives. TRACE-TS utilizes attribution from an expert classifier to pinpoint key spatio-temporal sensor areas, which are then employed to create directed acyclic graph (DAG) reasoning traces with clear evidence origins. A compact language model is trained to produce these traces through gated cross-attention on sensor memory tokens. During inference, the modified model simultaneously provides the activity prediction and its reasoning trace, removing the need for attribution at that stage. This framework aims to improve the dependability and traceability of AI-generated explanations in human activity comprehension, relevant to fields such as healthcare, smart environments, and human-computer interaction. The paper falls under artificial intelligence and signal processing and is part of the latest submissions on the arXiv preprint server.