Enhancing Fault Detection for Large Language Models via Mutation-Based Confidence Smoothing

Existing fault detection methods are not effective in detecting faults in large language models (LLMs), and a prompt mutation-based confidence smoothing method can significantly enhance the performance of these methods.