Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators

TL;DR

This paper introduces a decoding-time framework using comparators to reduce hallucinations in large language models, improving factual accuracy without altering internal model parameters.

Contribution

It proposes a novel CDT framework with instruction-guided comparators to enhance factuality during decoding, addressing hallucination patterns across tasks.

Findings

Significant improvement in response factuality across multiple tasks

Effective contrastive decoding constrains hallucination patterns

Comparator-driven decoding enhances model reliability

Abstract

Despite their remarkable capabilities, Large Language Models (LLMs) are prone to generate responses that contradict verifiable facts, i.e., unfaithful hallucination content. Existing efforts generally focus on optimizing model parameters or editing semantic representations, which compromise the internal factual knowledge of target LLMs. In addition, hallucinations typically exhibit multifaceted patterns in downstream tasks, limiting the model's holistic performance across tasks. In this paper, we propose a Comparator-driven Decoding-Time (CDT) framework to alleviate the response hallucination. Firstly, we construct hallucinatory and truthful comparators with multi-task fine-tuning samples. In this case, we present an instruction prototype-guided mixture of experts strategy to enhance the ability of the corresponding comparators to capture different hallucination or truthfulness patterns in distinct task instructions. CDT constrains next-token predictions to factuality-robust distributions by contrasting the logit differences between the target LLMs and these comparators. Systematic experiments on multiple downstream tasks show that our framework can significantly improve the model performance and response factuality.