When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence

Researchers defined derived-feature over-trust, a failure mode where an LLM treats a derived measurement as a direct fact or uses it outside its valid scope.

Using paired PPG and ECG physiological records, they introduced metrics for conflict, context errors, repair, evidence usefulness, and unnecessary verification.

An ECG-guided distillation baseline modestly improved several repair and specificity measures on a held-out test set.

The work provides a concrete benchmark-style framework for studying trust calibration when AI systems rely on imperfect model- or sensor-derived signals.