Can LLMs Introspect? A Reality Check

Shashwat Singh, Tal Linzen, Shauli Ravfogel

Can large language models detect and report their own internal states? A number of recent studies have argued that they can. Drawing on lessons from human metacognition research, w…