This paper reviews what researchers know so far about metacognition in large language models (LLMs). Metacognition means the ability to monitor and regulate one’s own thinking — for example judging confidence, choosing strategies, or deciding to seek more information. The authors describe a simple metacognitive loop made of monitoring (noticing uncertainty or progress) and control (planning and changing strategies based on that monitoring).

The team — drawing on work from Yale and the University of California, Irvine — presents what they call the first comprehensive overview of the field. They organize and classify existing studies, list benchmarks and measurement methods, and summarize techniques that try to elicit, improve, or apply metacognitive behavior in LLMs. The review also collects findings about when models show metacognitive-like behavior, and it points readers to an organized list of papers hosted on a public GitHub repository.

The authors explain why metacognition matters for AI. In humans, metacognition helps learning, problem solving, and decision making. For LLMs, the same abilities could boost task performance, make model outputs more trustworthy and interpretable, improve collaboration with people, and help reduce “hallucinations” (incorrect or made-up statements). The paper notes current interest because LLMs are already used in sensitive areas such as medicine, law, and scientific work.

The review reports mixed results from past studies. Some work finds that prompting models to reflect on their own answers can improve reasoning and make reported uncertainties more faithful. Other studies conclude that models are still unreliable at producing accurate confidence judgments. The authors highlight important differences between human metacognition and what LLMs currently do, and they flag many open questions about whether metacognitive skills can be trained, whether they generalize across tasks, and how model size or training methods affect them.