Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs

Tianyi Zhao, Yinhan He, Wendy Zheng, Yujie Zhang, Chen Chen · 2026

Read Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs by Tianyi Zhao, Yinhan He, Wendy Zheng, Yujie Zhang, Chen Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-01457.txt.

Description

Zhao and colleagues analyse high verbalized confidence in language models. It has three axes: tracing it as a gradient-friendly internal signal, finding the circuits that inflate it, and using these insights for test-time recalibration. Across two tuned language models on three test sets, a set of MLP blocks and attention heads, concentrated in middle-to-late layers, writes the confidence-inflation signal at the token position. Interventions on these circuits substantially improve calibration. Results suggest overconfidence is driven by identifiable circuits and can be mitigated through targeted intervention.

Catalog record

CREATORTianyi Zhao, Yinhan He, Wendy Zheng, Yujie Zhang, Chen Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42875
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.01457

Browse the full library · About the Tower of Babel