TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

Summary

TRACE-TS grounds explanations in attribution from an expert classifier — it finds the salient spatio-temporal sensor regions the classifier actually used, builds a structured reasoning trace (a DAG with explicit evidence provenance) from them, and trains a compact language model to reproduce that trace via gated cross-attention over sensor memory tokens, without needing the attribution computation at inference time.