DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations
This addresses the issue of unreliable outputs in LLMs for users requiring high factual accuracy, though it is an incremental improvement on existing decoding methods.
The paper tackles the problem of hallucinations in Large Language Models by proposing DeCoRe, a training-free decoding strategy that contrasts outputs from base and masked models to improve contextual faithfulness, resulting in performance gains such as 18.6% on XSum summarization and 10.9% on MemoTrap instruction following.
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%).