Cost-Effective Hallucination Detection for LLMs

TL;DR

This paper presents a cost-effective, multi-scoring framework for post-hoc hallucination detection in large language models, improving reliability while reducing computational costs across various tasks.

Contribution

It introduces a multi-scoring approach that combines different detection scores and a calibration method, enhancing performance and efficiency in hallucination detection.

Findings

Calibrating scores improves risk-aware decision making.

Multi-scoring outperforms individual scores across datasets.

Cost-effective methods match or surpass expensive detection techniques.

Abstract

Large language models (LLMs) can be prone to hallucinations - generating unreliable outputs that are unfaithful to their inputs, external facts or internally inconsistent. In this work, we address several challenges for post-hoc hallucination detection in production settings. Our pipeline for hallucination detection entails: first, producing a confidence score representing the likelihood that a generated answer is a hallucination; second, calibrating the score conditional on attributes of the inputs and candidate response; finally, performing detection by thresholding the calibrated score. We benchmark a variety of state-of-the-art scoring methods on different datasets, encompassing question answering, fact checking, and summarization tasks. We employ diverse LLMs to ensure a comprehensive assessment of performance. We show that calibrating individual scoring methods is critical for ensuring risk-aware downstream decision making. Based on findings that no individual score performs best in all situations, we propose a multi-scoring framework, which combines different scores and achieves top performance across all datasets. We further introduce cost-effective multi-scoring, which can match or even outperform more expensive detection methods, while significantly reducing computational overhead.