Beyond Yes and No: Improving Zero-Shot LLM Rankers via Scoring Fine-Grained Relevance Labels
Topics: AI (Deep Learning), Document Classification, Entity based search, Featured Snippets, Ranking, Retrieval Augmented Generation (RAG)
This research paper by Google presents a novel approach to improve zero-shot Large Language Model (LLM) document ranking systems by introducing fine-grained relevance labels instead of simple binary (yes/no) relevance judgments. The researchers from Google Research demonstrate that by allowing LLMs to score documents using multiple relevance levels (like “Highly Relevant,” “Somewhat Relevant,” and “Not Relevant”), the ranking performance significantly improves across multiple benchmark datasets. This improvement holds true even for datasets with binary ground-truth labels.