DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations

TL;DR

DeCoRe is a decoding strategy that reduces hallucinations in large language models by contrasting outputs from the original and masked models, improving factual accuracy in tasks like summarization and question answering.

Contribution

It introduces a training-free method that enhances LLM faithfulness by contrasting outputs of the base and masked models using conditional entropy.

Findings

Significant improvement in summarization accuracy (XSum +18.6%)

Enhanced instruction following (MemoTrap +10.9%)

Better open-book QA performance (NQ-Open +2.4%, NQ-Swap +5.5%)

Abstract

Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies have identified specific attention heads within the Transformer architecture, known as retrieval heads, responsible for extracting relevant contextual information. We hypothesise that masking these retrieval heads can induce hallucinations and that contrasting the outputs of the base LLM and the masked LLM can reduce hallucinations. To this end, we propose Decoding by Contrasting Retrieval Heads (DeCoRe), a novel training-free decoding strategy that amplifies information found in the context and model parameters. DeCoRe mitigates potentially hallucinated responses by dynamically contrasting the outputs of the base LLM and the masked LLM, using conditional entropy as a guide. Our extensive experiments confirm that DeCoRe significantly improves performance on tasks requiring high contextual faithfulness, such as summarisation (XSum by 18.6%), instruction following (MemoTrap by 10.9%), and open-book question answering (NQ-Open by 2.4% and NQ-Swap by 5.5%).