Research map / Large language models / Reasoning and chain of thought

LLMs Show No Signs Of Individuated Metacognition

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Decomposes binary confidence judgements from 20 frontier LLMs across six benchmarks via tetrachoric factor analysis and pairwise calibration, finding no significant individuated metacognition and attributing the MathBench exception to inline solving.

Problem

Whether LLMs' verbalised confidence reflects model-specific (individuated) metacognitive assessment of their own capabilities, beyond a shared item-difficulty signal, which is assumed by confidence-weighted routing, abstention and ensemble weighting.

Approach

Applies tetrachoric factor analysis to binary cross-model confidence matrices and pairwise Kendall tau-b calibration with base-rate-matched and calibration-preserving nulls to separate shared difficulty signal from model-specific calibration; also tests whether surface features of confidence-judgement reasoning text beat the model's own binary judgement.

Key results

• On SQuAD prospective, performance-factor and confidence-factor alignment is r=0.591 unfiltered but drops to r=0.230 on contentious items.
• Mean pairwise Kendall tau-b is small (e.g. SQuAD 0.037, OmniMath 0.025) and vanishes when restricted to similar-performance pairs; MathBench signal (tau=0.072) concentrates in reasoning-non-reasoning pairs (median tau=0.129).
• A ridge logistic regression on surface features of the confidence-judgement reasoning trace matches or beats the model's internal binary self-assessment AUC on three of six benchmarks (LegalBench, MathBench, MMLU).

Surprising findings

• Cross-model population consensus outperforms the individual model's own 20-sample yes-rate at predicting its own correctness in all but one near-identical case.
• On MathBench, apparent pairwise calibration is driven by reasoning models solving problems inline during the confidence probe rather than introspection.

Limitations

• Single binary yes/no judgement at temperature zero; masks variance on uncertain items.
• Reasoning/non-reasoning stratification relies on public documentation and may miscategorise borderline models.
• Tetrachoric latent-normality assumption is untested and likely violated, inflating dimensionality for extreme base rates.

Experimental setup

Datasets: SQuAD MMLU-Pro LegalBench MATH Omni-MATH SciCode

Benchmarks: SQuAD (modified) MMLU-Pro LegalBench MathBench Omni-MATH SciCode (modified)

Metrics: tetrachoric eigenvalues Kendall tau-b ROC AUC Pearson r F-beta

Baselines: Internal binary self-assessment cross-model population consensus sampled-yes-rate confidence

Builds on: tetrachoric correlation signal-detection metacognition (meta-d') Kendall tau-b parallel analysis

Closest papers by meaning

• Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.90)
• Decoupling Metacognition from Cognition: A Framework for Quantifying Metacognitive Ability in LLMs (AAAI 2025 · similarity 0.90)
• Tuning-Free Accountable Intervention for LLM Deployment – a Metacognitive Approach (AAAI 2025 · similarity 0.88)
• To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.87)
• Reasoning Models Better Express Their Confidence (NeurIPS 2025 · similarity 0.87)
• SteerConf: Steering LLMs for Confidence Elicitation (NeurIPS 2025 · similarity 0.87)