TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding
TRACE-TS introduces a framework for structured, signal-grounded reasoning over wearable sensor time series, addressing the weak grounding problem in existing LM-based approaches The method uses attribution from an expert classifier to identify salient spatio-temporal sensor regions and constructs DAG reasoning traces with explicit evidence provenance A compact language model is trained via gated cross-attention over sensor memory tokens to generate these reasoning traces, with inference requirin
Analysis
TL;DR
- TRACE-TS introduces a framework for structured, signal-grounded reasoning over wearable sensor time series, addressing the weak grounding problem in existing LM-based approaches
- The method uses attribution from an expert classifier to identify salient spatio-temporal sensor regions and constructs DAG reasoning traces with explicit evidence provenance
- A compact language model is trained via gated cross-attention over sensor memory tokens to generate these reasoning traces, with inference requiring no attribution computation or teacher guidance
- Semantic Node Match (SNM), an LLM-as-judge metric, is proposed to diagnose reasoning fidelity at observation, inference, and synthesis levels, localizing hallucinations missed by standard NLG metrics
- TRACE-TS achieves best average accuracy (84.43%) and F1 (81.24%) across seven wearable benchmarks, outperforming the best LLM-based baseline by 17.96% in F1
Why It Matters
This work addresses a critical gap in AI-driven human activity recognition: the tendency of language models to produce fluent but unverifiable explanations that lack grounding in actual sensor data. For AI practitioners building wearable-based health or behavioral monitoring systems, TRACE-TS provides a practical pathway to generate trustworthy, traceable reasoning that can be audited and validated against raw sensor signals.
Technical Details
- TRACE-TS employs attribution-based identification of salient spatio-temporal regions in wearable sensor data, using an expert classifier to pinpoint which sensor signals and time windows are most relevant to a given activity
- DAG (Directed Acyclic Graph) reasoning traces are constructed with explicit evidence provenance, creating structured chains of observation → inference → synthesis that can be audited
- A compact language model is trained through gated cross-attention mechanisms over sensor memory tokens, enabling joint output of activity predictions and reasoning traces at inference without requiring attribution computation or distillation from a teacher model
- Semantic Node Match (SNM) is an LLM-as-judge evaluation metric that operates at three fidelity levels—observation, inference, and synthesis—to detect hallucinated observations and broken evidence chains that standard NLG metrics fail to catch
- Evaluated across seven wearable benchmarks, achieving 84.43% accuracy and 81.24% F1, with a 17.96% F1 improvement over the best LLM-based baseline
Industry Insight
- The integration of attribution-based grounding with language model reasoning sets a new standard for explainable AI in wearable and IoT applications, where trust and verifiability are paramount for clinical or safety-critical deployments
- The SNM metric offers a reusable evaluation framework for any domain where LLM-generated reasoning must be grounded in structured data, potentially extending beyond wearable sensors to medical imaging, industrial monitoring, and autonomous systems
- The teacher-free inference design—eliminating the need for attribution computation at test time—makes TRACE-TS practically deployable in resource-constrained edge environments, bridging the gap between research prototypes and real-world wearable AI products
Disclaimer: The above content is generated by AI and is for reference only.