IBM's STAIR retriever cuts hallucination 65x vs fine-tuned RAG, hits 82.6% Recall@1

Promptmethus · x · 2026-10-28

IBM researchers unveiled STAIR (STructure Aware Information Retriever), an LLM-based retriever that stores a corpus's global structure (table-of-contents style) directly in model parameters instead of chunking by length and relying on vector similarity.

Core claim: standard RAG chunks discard semantic global structure, and when the retriever pulls irrelevant documents the model just "hallucinates with citations."
Results: below 0.05% hallucination in a generative IR system — 65x less than fine-tuned DSI systems; 82.6% Recall@1 on their new SearchTome benchmark (18 books, 6 domains) vs DSI 76.9%, DPR 68.7%, BM25 59.5%, Mistral 13.8%.
Code and benchmark released on GitHub and Hugging Face.

Original post →

IBM's STAIR retriever cuts hallucination 65x vs fine-tuned RAG, hits 82.6% Recall@1

More from Models