Internal Flow Signatures for Self-Checking and Refinement in LLMs

Sungheon Jeong, Sanggeon Yun, Ryozo Masukawa +3

Large language models can generate fluent answers that are unfaithful to the provided context, while many safeguards rely on external verification or a separate judge after generat…