Stable Miscalibration in Large Language Models: A Practical View of High-Confidence Errors

Akira Okutomi

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 confid…