Large language models reduce confident wrong answers by fixing internals
Why Are LLMs Confidently Wrong? Correcting Overconfident Errors via Causal Head Intervention
Artificial Intelligence
Summary
Sometimes large language models, or LLMs, are very sure about their answers even when they are wrong, which can cause problems when people rely on them. The authors show that most ways to fix this only adjust the model's final confidence scores after the fact, without understanding why the errors happen inside. They propose a new method called CoCHI that looks at specific parts inside the model responsible for how confident it feels, then changes those parts to make it less confidently wrong. Tests on several models and quizzes show CoCHI makes the models’ confidence more honest without hurting how often they get answers right.
What this means in practice
- •For ai product developers: Improve trustworthiness of large language model outputs by reducing overconfident incorrect answers through internal adjustments.
- •For customer support teams: Deploy chatbots that better express uncertainty to avoid misleading confident but wrong responses in automated customer interactions.
Authors
Jing Ren, Bowen Li, Ziqi Xu, Xuechao Yang, Feng Xia
Abstract
Modern large language models (LLMs) often exhibit overconfident errors, assigning high confidence to incorrect predictions and thereby undermining reliability in real-world use. Existing calibration methods typically rely on post-hoc adjustment or retraining, operating at the output level without addressing the underlying causes of miscalibration. We propose CoCHI, (Class-oriented Causal Head Intervention), a framework that improves calibration by identifying and intervening on internal components responsible for distinct confidence behaviours. By analysing predictions through correctness-confidence patterns, CoCHI enables targeted intervention on model internals, reducing overconfident errors while preserving correct predictions. Our approach provides a mechanistic perspective on calibration, linking reliability to identifiable structures within the model rather than treating it as an output-level issue. Experiments across multiple LLMs show consistent improvements in calibration metrics, including Expected Calibration Error, Negative Log-Likelihood, and Brier Score, while maintaining accuracy on five multiple-choice question answering benchmarks. These results suggest that effective calibration requires not only adjusting outputs, but also addressing the underlying causal mechanisms within the model.