A framework has been introduced to improve confidence estimation in Large Language Models (LLMs). It integrates a structured reasoning process with a distance-aware calibration technique.
The approach features three key innovations: a Maximum Confidence Selection (MCS) method that assesses confidence across all labels, a reflection-based prompting mechanism to improve reasoning reliability, and a calibration technique that considers ordinal label relationships.
Evaluation on datasets such as HelpSteer2, Llama T-REx, and a proprietary conversational dataset demonstrates the framework's effectiveness in both conversational and fact-based classification tasks. This work aims to support more reliable confidence estimation, aiding decision-making processes in model deployment.
Source: https://arxiv.org/abs/2609.04539