Reconfidencing LLMs from the Grouping Loss Perspective

TL;DR

A solution to reconfidence LLMs is proposed, canceling not only calibration but also grouping loss, and the LLMs, after the reconfidencing process, indicate improved confidence alignment with the accuracy of their responses.

Abstract

Large Language Models (LLMs), including ChatGPT and LLaMA, are susceptible to generating hallucinated answers in a confident tone. While efforts to elicit and calibrate confidence scores have proven useful, recent findings show that controlling uncertainty must go beyond calibration: predicted scores may deviate significantly from the actual posterior probabilities due to the impact of grouping loss. In this work, we construct a new evaluation dataset derived from a knowledge base to assess confidence scores given to answers of Mistral and LLaMA. Experiments show that they tend to be overconfident. Further, we show that they are more overconfident on some answers than others, \emph{eg} depending on the nationality of the person in the query. In uncertainty-quantification theory, this is grouping loss. To address this, we propose a solution to reconfidence LLMs, canceling not only calibration but also grouping loss. The LLMs, after the reconfidencing process, indicate improved confidence alignment with the accuracy of their responses.

Figures (11)

• Figure 1: Desired user experience -- An illustration of our goals of eliciting confidence levels in LLMs. High confidence scores are represented in green, while red indicates a higher likelihood of encountering hallucinated sentences.
• Figure 2: Calibration curves of the Birth_Date relation. The LLM here is Mistral-7B jiang2023mistral, and we use SelfCheckGPT manakul2023selfcheckgpt to compute confidence scores. An increased number of $\star$ symbols signifies a sub-group containing more popular samples.
• Figure 3: Grouping diagrams of latent sub-groups. These groups are created from the leaves of a decision tree. SCGPT is an abbreviation for SelfCheckGPT.
• Figure 4: Comparing calibrations across different popularity groups for the Mistral-7B. We use merged results of three regions. The confidence method here is SelfCheckGPT. More $\star$ symbols mean a sub-group with more popular samples.
• Figure A1: Grouping diagrams of user-defined sub-groups. We divide each bin into eight groups by the popularity of entities. SCGPT is an abbreviation for SelfCheckGPT.