DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations

Large Language Models (LLMs) often hallucinate by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies identified specific attention heads that are responsible for extracting relevant contextual information, known as retrieval heads. We hypothesise that masking these heads can induce hallucinations and that we can reduce hallucinations by contrasting the outputs of the base and the masked LLMs. 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 and the masked LLMs, using conditional entropy as a guide. Our results show that DeCoRe significantly improves performance on tasks requiring high contextual faithfulness, such as summarisation (+18.6% on XSum), instruction following (+10.9% on MemoTrap), and openbook question answering (+2.4% on NQ-Open and +5.5% on NQ-Swap). 1