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

Researchers have introduced a method to curb overconfidence in large language models without undercutting their ability to express confidence when correct, according to a paper posted to arXiv on June 2 [1]. The technique, called Probe-Conditioned Head Intervention (PCHI), operates at inference time. It uses a frozen probe to detect responses where the model is likely wrong but confident, then conditionally rescales the outputs of downstream attention heads during confidence generation [1]. Standard calibration methods often act globally, which can reduce unwarranted confidence but also erode warranted confidence on correct answers [1]. PCHI is designed to avoid that trade-off by intervening only when the probe flags a high risk of a wrong-yes response [1]. On the Qwen3-4B-Instruct model solving OpenMathInstruct problems with a structured binary confidence field, readout-token PCHI converted 82.2% of originally wrong-yes confidence readouts to “no” [1]. A joint intervention across upstream confidence-template tokens reduced the Expected Calibration Error (ECE) from 21.9% to 9.2% while damaging only 5.1% of originally correct-yes readouts [1]. The readout-token effect also appeared on Gemma3-4B, though the upstream interventions were weaker and more dependent on the specific mask used [1]. The work addresses a known limitation of large language models: they frequently express high confidence in incorrect answers [1]. By selectively suppressing verbalized overconfidence through conditionally applied internal intervention, the method partially decouples the suppression of unwarranted confidence from the loss of warranted confidence [1]. The paper does not include quotes from the authors. PCHI’s conditional approach differs from prior calibration techniques that apply uniform adjustments across all model outputs. The probe acts as a gate, leaving correct-yes responses largely intact while redirecting wrong-yes responses toward a “no” confidence readout [1]. The authors report that the intervention is applied only during confidence generation, leaving the underlying answer-generation process unchanged [1].