Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors
The paper discusses the phenomenon of Authority Inversion in large language models (LLMs) used in ubiquitous systems. It highlights how these models may prioritize user claims over sensor data, raising concerns about reliability. The authors propose a geometric framework and intervention methods to address this issue and improve decision-making accuracy.
- ▪Large language models increasingly integrate various inputs, but their authority allocation when user claims conflict with sensor data is not well understood.
- ▪The study reveals that LLMs often exhibit extreme authority inversion, where numerical sensor data is largely disregarded in favor of user claims.
- ▪The authors introduce metrics and a calibration method to improve the reliability of LLMs in decision-making scenarios.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.23938 (cs) [Submitted on 28 Apr 2026] Title:Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors Authors:Long Zhang, Zi-bo Qin, Wei-neng Chen View a PDF of the paper titled Authority Inversion in LLM-Mediated Ubiquitous Systems: When Models Trust Users Over Sensors, by Long Zhang and 2 other authors View PDF HTML (experimental) Abstract:Large language models (LLMs) increasingly fuse heterogeneous inputs in ubiquitous systems. Yet, how LLMs implicitly allocate authority when sensor measurements and user claims conflict remains unexamined, raising critical reliability concerns for deployments where physical sensing must retain priority.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
