Summary of Decore: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations, By Aryo Pradipta Gema et al.
DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations
by Aryo Pradipta Gema, Chen Jin, Ahmed Abdulaal, Tom Diethe, Philip Teare, Beatrice Alex, Pasquale Minervini, Amrutha Saseendran
First submitted to arxiv on: 24 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This research paper proposes a novel training-free decoding strategy called Decoding by Contrasting Retrieval Heads (DeCoRe) to reduce hallucinations in Large Language Models (LLMs). The authors identify specific attention heads, known as retrieval heads, responsible for extracting contextual information and hypothesize that masking these heads can induce hallucinations. DeCoRe amplifies contextually faithful responses by dynamically contrasting the outputs of a base LLM with a masked LLM guided by conditional entropy. Experimental results show significant improvements in tasks requiring high contextual faithfulness, such as summarization (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%). |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research helps make language models more accurate by reducing mistakes that happen when they try to understand the context of what’s being asked. The scientists found a way to stop these “hallucinations” from happening in the first place. They did this by changing how the model gets its information, making it rely on the actual context instead of just guessing. This new method works really well and can help improve things like summarizing text or answering questions correctly. |
Keywords
» Artificial intelligence » Attention » Question answering » Summarization
