This paper provides the first comprehensive review of metacognition for large language models, analyzing and taxonomizing the current landscape of this emerging field.

  • Methods and benchmarks to measure and evaluate LLMs' metacognitive abilities are summarized.
  • Techniques to elicit, improve, and apply metacognition in LLMs are detailed.
  • Findings, implications, open questions, and challenges of ongoing research are discussed.

The authors aim to provide a detailed overview to stimulate meaningful research and discussion on advancing AI capabilities and reliability.