Prompting Language Models Improves Quoting from Pre-Training Data - EACL 2024

Explore a 10-minute conference talk presented by Marc Marone at EACL 2024, discussing the paper "According to …: Prompting Language Models Improves Quoting from Pre-Training Data." Delve into the innovative "according-to prompting" technique, designed to enhance the factual accuracy of Large Language Models (LLMs) by grounding their responses in previously observed text. Learn about the novel QUIP-Score evaluation metric, which measures how well model-generated answers align with underlying text corpora. Examine experiments conducted on Wikipedia, PubMed, and U.S. legal tax code, demonstrating improved grounding and task performance. Discover how LLMs can increase or decrease grounded generations on request, offering potential solutions to combat hallucination and fake information generation in AI language models.