Metacognition

Metacognition Slug: metacognition

Metacognition refers to the capacity to monitor, evaluate, and regulate one's own cognitive processes: knowing what one knows, knowing what one does not know, judging the reliability of one's own inferences, and selecting strategies accordingly. The concept originates in cognitive psychology and developmental psychology (Flavell 1979; Nelson & Narens 1990), where it describes how human learners track their own learning state and adapt study strategies, allocation of effort, and information-seeking behavior.

Applied to large language models, metacognition covers four interacting capabilities: (1) self-knowledge, which means recognizing the model's knowledge boundaries and refusing to answer when uncertain rather than hallucinating; (2) uncertainty calibration, which means mapping internal confidence (logit distributions, ensembling, or self-consistency) into numerically accurate probabilities; (3) strategy selection, which means choosing which prompting, decoding, or tool-use strategy to apply based on the perceived difficulty of the question; and (4) self-explanation, which means producing natural-language reports about its own reasoning and confidence that are faithful to the underlying computation. Frontier LLMs exhibit partial but uneven metacognitive abilities. They often hallucinate with high confidence, fail to recognize knowledge boundaries, and misrepresent their internal uncertainty.

Metacognition in LLMs is studied under several subheadings: metacognitive monitoring (Lin 2025, "Language Models Are Capable of Metacognitive Monitoring"), metacognitive data selection (Liu et al. 2026, "Reinforcement Learning with Metacognitive Feedback"), and uncertainty communication (Maheshwari 2025, "Metacognition and Uncertainty Communication in Humans and LLMs"). A 2026 survey (Wang et al., "Metacognition in LLMs: Foundations, Progress, and Opportunities") consolidates the area. Metacognition is adjacent to self-awareness (the broader capacity for self-monitoring), uncertainty estimation (the numerical sub-task), and faithful calibration (the alignment of expressed with intrinsic uncertainty).