A new study reveals that large language models (LLMs) often prioritize user claims over sensor data, leading to reliability issues in decision-making. The authors propose a framework and intervention method to better align authority allocation, significantly improving model accuracy in numerical tasks.