Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors
A new arXiv paper (2605.23938) from researchers studying LLM-mediated ubiquitous systems uncovers a critical flaw: when sensor measurements and user claims conflict, LLMs implicitly allocate authority in a severely format-dependent manner. Numerical sensor data fails to integrate into answer-relevant model directions, allowing natural-language user claims to dominate final decisions—a phenomenon termed 'Authority Inversion.' To diagnose this, the authors develop a geometric framework of context integration and introduce two audit metrics: Context Integration Ratio (CIR) and Authority Alignment Index (AAI). They propose Geometric Authority Calibration (GAC), an inference-time layer-level intervention to suppress misplaced user authority. Evaluating four models (4B to 35B parameters, three architectures) across four datasets totaling 576 conflict instances reveals extreme inversion. The paper is published on arXiv and has not yet been peer-reviewed.
Developers must ensure sensor data is properly formatted to avoid LLM overriding physical measurements.