Reproduce “DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aryo Pradipta Gema, Chen Jin, Ahmed Abdulaal, Tom Diethe, Philip Teare, Beatrice Alex, Pasquale Minervini, Amrutha Saseendran
A novel training-free decoding strategy, DECORE, addresses LLM hallucinations by contrasting the base model's output with a variant where identified "retrieval heads" are masked. This approach leads to enhanced contextual faithfulness, with an 18.6% improvement in factKB score on XSum, better factual recall, and stronger multi-hop reasoning capabilities.
Key insights
- Specific "retrieval heads" within an LLM are mechanistically responsible for extracting and utilizing contextual information, and their impairment directly induces hallucinations.
- Contrasting the outputs of a base LLM with a variant where its retrieval heads are disabled provides a targeted, training-free method to suppress hallucinated generations.
- Dynamically adjusting the contrastive decoding strength based on the LLM's conditional entropy effectively guides the model away from uncertain and potentially unfaithful outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.