Estimating LLM Uncertainty with Evidence
This work addresses the issue of unreliable responses in LLMs for users in various domains, but it is incremental as it builds on existing uncertainty estimation methods by focusing on evidence strength.
The paper tackles the problem of hallucinations in Large Language Models by addressing the poor performance of probability-based methods in estimating token reliability, revealing that this is due to lost evidence strength information. It introduces Logits-induced token uncertainty (LogTokU), a framework that enables real-time uncertainty estimation without multiple sampling, and experimental results show it has significant effectiveness.
Over the past few years, Large Language Models (LLMs) have developed rapidly and are widely applied in various domains. However, LLMs face the issue of hallucinations, generating responses that may be unreliable when the models lack relevant knowledge. To be aware of potential hallucinations, uncertainty estimation methods have been introduced, and most of them have confirmed that reliability lies in critical tokens. However, probability-based methods perform poorly in identifying token reliability, limiting their practical utility. In this paper, we reveal that the probability-based method fails to estimate token reliability due to the loss of evidence strength information which is accumulated in the training stage. Therefore, we present Logits-induced token uncertainty (LogTokU), a framework for estimating decoupled token uncertainty in LLMs, enabling real-time uncertainty estimation without requiring multiple sampling processes. We employ evidence modeling to implement LogTokU and use the estimated uncertainty to guide downstream tasks. The experimental results demonstrate that LogTokU has significant effectiveness and promise.