STAIR Retrieves Through a Table of Contents Instead of Length-Based Chunks and Holds Hallucination Under 0.05%

arXiv 2609.03874·medium signalXBlueskyLinkedInCopy link

Current retrievers chop long context into length-based chunks and throw away the corpus's global semantic structure, which still matters because long-context models keep losing information in the middle. STAIR lets an LLM exploit that global structure, such as a table of contents, to store and retrieve from its own parameters, and ablations against a finetuned Differentiable Search Index show ToC grounding yields a generative IR system with under 0.05% hallucination that generalizes when training samples are scarce. STAIR reaches Recall@1 of 82.6% on the released SearchTome benchmark against DSI's 76.9%, a statistically significant gap; SearchTome itself spans 18 books across 6 domains and is released for follow-on work.