Do AI know what they know? Exploring metacognition in LLMs

Large Language Models (LLMs) have demonstrated remarkable proficiency in natural language processing applications, encompassing question answering, text generation, and reasoning capabilities. However, their metacognitive abilities, which involve self-assessment, uncertainty awareness, and cognitive control, remain insufficiently explored. This investigation examines over 97 publications released between 2021 and 2025. These studies encompass prompt-based methodologies, fine-tuning techniques, retrieval-augmented generation, and agentic AI frameworks that implement metacognition within lifelong learning models (LLMs). Metacognitive interventions produce measurable, task-dependent performance gains. Reported improvements range from approximately 3% to over 20% across evaluated benchmarks. These gains are observed in medical reasoning, multi-hop question answering, and natural language comprehension tasks. These enhancements correlate with demonstrable improvements in confidence calibration, error detection, and the precision of response revisions. Furthermore, this work conducts a comprehensive examination of more than ten families of large language models and agentic frameworks, systematically identifying enduring challenges, including overconfidence, susceptibility to hallucinations, limited error awareness, and unstable self-reflection processes. This research furnishes an analytical basis for elucidating the conditions under which metacognitive mechanisms either enhance or diminish LLM reliability. Moreover, it delineates prospective research avenues aimed at developing scalable, trustworthy, and human-centered artificial intelligence systems. This foundation is established through the synthesis of evaluation protocols, benchmarking methodologies, and comparative evidence.