Pre-trained Language Models Return Distinguishable Probability Distributions to Unfaithfully Hallucinated Texts
This addresses the problem of text hallucination in language models for users relying on accurate AI-generated content, offering a method to enhance faithfulness, though it appears incremental as it builds on known model behaviors.
The paper demonstrates that pre-trained language models produce statistically distinguishable probability and uncertainty distributions for unfaithfully hallucinated texts across 24 models and 6 datasets, with 88-98% of cases showing this pattern. It introduces a hallucination-reducing training algorithm that improves faithfulness metrics while preserving general text quality.
In this work, we show the pre-trained language models return distinguishable generation probability and uncertainty distribution to unfaithfully hallucinated texts, regardless of their size and structure. By examining 24 models on 6 data sets, we find out that 88-98% of cases return statistically significantly distinguishable generation probability and uncertainty distributions. Using this general phenomenon, we showcase a hallucination-reducing training algorithm. Our algorithm outperforms other baselines by achieving higher faithfulness metrics while maintaining sound general text quality measures.