A retrieval system that uses a book's table of contents to cut AI hallucinations below 0.05%.

When large language models search through long texts to answer questions, they often struggle with a "lost in the middle" issue. Standard search systems chop long documents into simple, length-based chunks. This cuts up the text and throws away the overall structure. To solve this, researchers built STAIR. STAIR trains a language model to use global outlines, like a Table of Contents. By learning this outline, the model can store and look up information directly inside its own parameters. This approach creates a system that hallucinates less than 0.05% of the time, and it works well even when very few training examples exist. The team tested it on SearchTome, a new benchmark made of 18 books across 6 diverse domains. STAIR scored 82.6% on Recall@1, finding the right information on its first try far better than standard search methods.

Why it matters

For anyone relying on AI to answer questions from long documents, this means fewer wrong answers and more trust in the information. It helps in fields like research, education, and customer support where accuracy matters.

Questions answered