Researchers discover that LLMs express confidence levels disconnected from actual accuracy due to orthogonal encoding of calibration and confidence signals.

Analysis using linear probes and contrastive activation addition (CAA) reveals that calibration and verbalized confidence are encoded linearly but operate independently across three open-weight models and four datasets.

A new 'Reasoning Contamination Effect' shows that when models simultaneously reason through problems and express confidence, the reasoning process disrupts confidence signals and worsens miscalibration.

Researchers propose a two-stage adaptive steering pipeline that reads the model's internal accuracy estimates to bridge the confidence-faithfulness gap in LLMs.