Can language models genuinely monitor their own thinking?

The first comprehensive survey of metacognition in LLMs — the ability to monitor, assess, and regulate one's own cognitive processes — closes on the question that actually matters for deployment: are models capable of genuine metacognition, or are they simulating memorized patterns? This is not a philosophical footnote. If a model's self-assessment is a learned surface behavior rather than a real read on its own states, then confidence reports, self-correction, and introspective explanations are unreliable exactly where they are needed. The survey ties this directly to safe oversight, deployment, and human-AI interaction, because oversight regimes increasingly lean on models reporting their own uncertainty.

The tension is that empirical evidence points both ways. Can language models detect their own internal anomalies? shows models detecting injected concepts before those perturbations shape output — evidence for something more than pattern-matching — while Can LLM explanations actually help humans predict model behavior? shows self-explanations that do not track what the model would actually do. Both can be true if metacognition is real but shallow and unevenly distributed across tasks. The survey's contribution is to make this a measurable research program rather than an assumption, taxonomizing methods to elicit, improve, and evaluate metacognitive ability. The practical upshot: any oversight design that trusts a model's self-report inherits an unresolved empirical bet, and that bet should be tested per capability, not granted wholesale.