Do LLMs estimate uncertainty well in instruction-following?

Motivation Methods Results P la c e h o ld e r Instruction-following matters for building reliable LLM agent. Deployed models must strictly follow the instructions and constraints from users to ensure that the outputs are both safe and aligned with user intentions. Since LLMs are prone to errors, uncertainty estimation ability in instruction-following is essential. If the LLM misinterprets or deviates from these instructions but accurately recognizes and signals high uncertainty, it could prompt further review or intervention, thereby preventing the delivery of potentially harmful advice. However, uncertainty estimation in instruction following tasks has received limited attention.