LLMs prioritize user claims over sensor data, researchers find

A new research paper identifies a critical issue in how large language models (LLMs) handle conflicting information from sensors and user claims in ubiquitous systems. The study, titled "Authority Inversion in LLM-Mediated Ubiquitous Systems," reveals that LLMs tend to prioritize natural-language user statements over numerical sensor data, a phenomenon termed "Authority Inversion." Researchers developed metrics like CIR and AAI to quantify this, finding models exhibit near-zero trust in sensor data. They also proposed Geometric Authority Calibration (GAC) to mitigate this by explicitly configuring authority allocation, significantly improving accuracy in tasks like human activity recognition. AI

影响 Highlights a critical flaw in LLM decision-making for real-world systems, necessitating explicit configuration for reliable sensor data integration.

排序理由 The cluster contains an academic paper detailing a novel finding about LLM behavior and proposing a mitigation strategy. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• Authority Alignment Index (AAI)
• Authority Inversion
• Context Integration Ratio (CIR)
• Geometric Authority Calibration (GAC)
• LLMs

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →