Masked by Consensus: Disentangling Privileged Knowledge in LLM Correctness

Tomer Ashuach, Shai Gretz, Yoav Katz +2

Humans use introspection to evaluate their understanding through private internal states inaccessible to external observers. We investigate whether large language models possess si…