On the Limits of Metacognitive Monitoring in LLMs
Abstract
Reliable decisions depend on recognizing when an answer may be wrong. In biological cognition, metacognitive monitoring can dissociate from task performance, raising the question of how closely solving and judging are linked in language models. Here we study the confidence reports of four frontier models across 15 benchmarks. High task accuracy can coexist with weak error discrimination: a model solves 97% of competition mathematics problems while its answer-time confidence ranks correct answers above errors barely better than chance. Confidence separates correct answers from errors more effectively on questions solved by a separate reference model, while review brings limited improvement on reference-hard questions. Aggregate discrimination also rewards ranking correct answers on easy questions above errors on hard ones, which question-only forecasts already do well. Cross-evaluation helps most where the evaluator answered correctly, and errors shared by the two models usually retain high confidence. Hard questions and shared errors remain difficult targets for prompted self-review and peer oversight, even in models with strong problem-solving performance.
Related
- Metacognitive Skill Learning as a Mechanism of Advanced Meditation
- Endogenous Attention Effects on Metacognitive Awareness: A Proof-of-Concept DLPFC-cTBS Study
- Neural Mechanisms Underlying Metacognitive Processes in Sequential Decision-Making
- Concurrent metacognitive evaluation is dissociable from perceptual evidence accumulation in the parietal cortex
- Developing Midlife Metacognitive Strategy Training (M-MCST) program: A feasibility study with middle-aged adults to evaluate the design, implementation, and impact of a novel program in fostering healthy cognitive aging
BibTeX
@misc{han2026limits,
title = {{On the Limits of Metacognitive Monitoring in LLMs}},
author = {Dongqi Han and Yifan Yang and Dongsheng Li},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.34864},
eprint = {2609.34864},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.34864},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - On the Limits of Metacognitive Monitoring in LLMs
AU - Dongqi Han
AU - Yifan Yang
AU - Dongsheng Li
PY - 2026
DA - 2026/09/28
PB - arXiv
DO - 10.48550/arxiv.2609.34864
UR - https://arxiv.org/abs/2609.34864
AB - Reliable decisions depend on recognizing when an answer may be wrong. In biological cognition, metacognitive monitoring can dissociate from task performance, raising the question of how closely solving and judging are linked in language models. Here we study the confidence reports of four frontier models across 15 benchmarks. High task accuracy can coexist with weak error discrimination: a model solves 97% of competition mathematics problems while its answer-time confidence ranks correct answers above errors barely better than chance. Confidence separates correct answers from errors more effectively on questions solved by a separate reference model, while review brings limited improvement on reference-hard questions. Aggregate discrimination also rewards ranking correct answers on easy questions above errors on hard ones, which question-only forecasts already do well. Cross-evaluation helps most where the evaluator answered correctly, and errors shared by the two models usually retain high confidence. Hard questions and shared errors remain difficult targets for prompted self-review and peer oversight, even in models with strong problem-solving performance.
ER -