Tag
This paper investigates how LLMs degrade in detecting planted document contaminants as batch size increases, leading to confident hallucinations of non-existent errors, and recommends bounded batch sizes and verification mechanisms for reliable auditing.