The Metacognitive Probe: Five Behavioural Calibration Diagnostics for LLMs

Researchers introduce the Metacognitive Probe, a diagnostic tool measuring five dimensions of LLM confidence behavior including calibration, epistemic vigilance, and reasoning validation. Testing on eight frontier models and 69 humans reveals significant within-model disparities—exemplified by Gemini 2.5 Flash scoring 88 on confidence calibration but only 41 on difficulty prediction—suggesting composite benchmarks mask pockets of overconfidence.

The Metacognitive Probe addresses a critical blind spot in current AI evaluation methodologies. While benchmarks like MMLU and HELM measure whether models produce correct answers, they remain silent on whether models accurately assess their own uncertainty. This distinction matters substantially: a model achieving 80% on standard benchmarks could harbor severe overconfidence in specific domains, creating dangerous failure modes in production environments where calibrated uncertainty is essential for user trust and safe deployment.

The research builds on foundational metacognition theory (Flavell, Nelson and Narens) but operationalizes it through observable confidence-correctness alignment rather than attempting cross-species psychological validation. The five-task framework decomposes model behavior into distinct dimensions—confidence calibration, epistemic vigilance (knowing when to doubt), knowledge boundary recognition, calibration range consistency, and reasoning-chain validation. This decomposition enables granular diagnostics impossible with aggregate metrics.

The Gemini 2.5 Flash case study illustrates the probe's value: a 47-point spread between panel-best calibration performance (T1-CC = 88) and panel-worst difficulty prediction (T4-CR = 41) suggests the model exhibits inconsistent self-awareness across task types. This finding has immediate implications for developers integrating frontier models into production systems where localized confidence failures could propagate harmful outputs despite strong average performance.

Future validation should establish whether these five dimensions represent exhaustive behavioral decomposition and whether the metrics correlate with real-world failure modes in deployed applications.

• →Composite benchmarks fail to detect pockets of extreme overconfidence that the Metacognitive Probe can surface through five distinct behavioral dimensions.
• →Gemini 2.5 Flash shows a 47-point performance gap between its best calibration task and worst difficulty prediction, indicating inconsistent self-awareness.
• →Current evaluation frameworks are blind to whether models know when their responses are wrong, a critical safety consideration for deployment.
• →The probe's decomposition into five behaviorally-distinct dimensions enables granular diagnostics impossible with aggregate metrics.
• →The research falsified the pre-specified human developmental hypothesis, suggesting metacognitive behavior in LLMs may not follow human developmental patterns.