Large language models (LLMs) can produce seemingly credible yet inaccurate responses, prompting researchers to enhance uncertainty quantification methods for reliability checks. A common technique involves repeated prompts to measure self-confidence, which may not truly reflect accuracy and can lead to significant risks in critical fields like healthcare and finance. MIT researchers have developed a novel method assessing uncertainty by analyzing responses from various similar LLMs, focusing on cross-model disagreement to gauge epistemic uncertainty—effectively identifying confident but incorrect outputs. This method combines cross-model analysis with self-consistency measures to generate a total uncertainty metric (TU), demonstrating superior performance across ten tasks including question-answering and math reasoning. Unlike traditional measures, TU can more reliably flag hallucinations in model outputs while requiring fewer queries, thereby reducing computational costs. Future research aims to adapt this method for open-ended tasks and further explore other forms of uncertainty, funded by the MIT-IBM Watson AI Lab.
Source link
Share
Read more