Calibrating Overconfidence Without Sacrificing Confidence: Probe-Conditioned Head Intervention for LLMs

In three linesPCHI, an inference-time method, detects wrong-but-confident responses using a frozen probe and conditionally rescales attention-head outputs. On Qwen3-4B and Gemma3-4B, it converts 82.2% of false confident answers to 'no' and reduces ECE from 21.9% to 9.2% while preserving 94.9% of correct answers.

Summary generated by Claude — human-verified