Stable Miscalibration in Large Language Models: A Practical View of High-Confidence Errors
06:00 · August 17, 2026 · arXiv cs.AI RSS
High-confidence errors in large language models are often treated as evidence of fragile internal inference. We study a different possibility: stable miscalibration, where a confident wrong answer remains locally stable under small perturbations. We combine two diagnostics: a label-aware output-level audit score that ranks domains by confidence variation and overconfident mistakes under a forced-answer baseline, and an internal sensitivity probe that measures hidden-state movement. On a multi-domain binary factual audit set, this audit score tracks where abstention-aware self-critique reduces decision loss, although direct labeled baselines rank the same gain more strongly. Internally, self-critical prompting consistently reduces hidden-state sensitivity across layers in three open-weight models. This supports prompt-induced local stabilization rather than a purely output-level abstention pattern, but it does not imply calibration: audit-defined overconfident errors are not clearly more locally sensitive than confidently correct answers, so some high-confidence errors may be stable and miscalibrated rather than simply fragile.
Summary
High-confidence errors in large language models are commonly viewed as signs of fragile internal inference, where small input changes readily flip an incorrect output. This paper examines an alternative: stable miscalibration, in which a confidently wrong answer persists under modest perturbations because the model remains near the same internal state. The distinction matters for uncertainty estimation, because fragility-based probes would then fail to separate overconfident mistakes from confidently correct answers.
To test the idea, the author combines two diagnostics on a frozen set of 532 binary factual items spanning eleven domains. An output-level audit score ranks domains according to confidence variation across three policies—a forced-answer baseline, a cautious abstention policy, and a self-critical abstention policy—plus the mass of overconfident mistakes produced by the forced-answer baseline. An internal sensitivity probe tracks layer-wise movement of hidden states under the same policies. Experiments cover three open-weight models, including Llama-3.1 and DeepSeek-R1.
The audit score identifies domains where abstention-aware self-critique reduces policy-aware decision loss, although direct labeled baselines capture the same improvement more strongly. Internally, self-critical prompting lowers hidden-state sensitivity across layers, indicating prompt-induced local stabilization rather than a purely output-level abstention effect. However, items that the audit flags as overconfidently wrong do not exhibit greater local sensitivity than confidently correct items. This pattern suggests that some high-confidence errors are locally stable yet miscalibrated, rather than merely fragile.
The findings imply that abstention policies and uncertainty estimates should account for stable miscalibration in addition to perturbation sensitivity when deciding when a model should refrain from answering.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners focused on trustworthy and transparent AI, a key priority in the Netherlands and the EU. Understanding stable miscalibration provides actionable insights for improving LLM reliability, auditing high-confidence hallucinations, and developing better abstention-aware decision policies for enterprise deployment.
