IBM's STAIR uses table-of-contents retrieval to cut IR hallucinations to under 0.05%
alex_verem · x · 2026-10-14
IBM researchers introduced STAIR (STructure Aware Information Retriever), a generative retrieval system that exploits global document structure (tables of contents) instead of vector databases, embedding search, or re-rankers.
- Core idea: Standard RAG chunks long documents by length, discarding rich global semantic structure. STAIR lets an LLM store and retrieve information via the ToC — like a human skimming a 900-page textbook by chapter — avoiding the lost-in-the-middle problem.
- Results: Against a fine-tuned Differentiable Search Index (DSI) baseline, the ToC-based system achieves under 0.05% hallucination (65x fewer than its closest rival) and generalizes with very few training samples.
- Benchmark: The team also released SearchTome, a benchmark built from 18 books across 6 domains, where STAIR hits 82.6% Recall@1.
The paper with full ablations is on arXiv (2609.03874).
Related event: IBM's STAIR uses document TOCs to cut hallucination rate to 0.05%(2 posts)→
More from Research
- Submarine cyber ops study: operator cognitive capacity swings with the tactical environment — J_MaestreVidal · 2026-10-14
- Recova Trains Robots to Self-Recover in Simulation, Cutting Human Intervention From 87.5% to 0% — Lianhuiq · 2026-10-14
- NVIDIA's AlphaNIST Learns Robotic Assembly From Scratch With Only Binary Rewards, Hitting 98.2% Across 20 Tasks — abhishekunique7 · 2026-10-14
- Encrypted Chain-of-Thought Can Be Decoded: Reasoning Theft Attack on Top LLM APIs — matthew_d_green · 2026-10-14
- Most people carry their mother's cells in their brain, even at 80 — AnnaCiaunica · 2026-10-14
- Looping Layers Boosts Small Decision Models: 1.4B Nears 3x Larger Model — rohanpaul_ai · 2026-10-14