IBM's STAIR retriever cuts hallucinations 65x by using table of contents instead of chunking
alex_verem · x · 2026-11-03
IBM researchers introduced STAIR (STructure Aware Information Retriever), challenging the standard RAG practice of fixed-length chunking, which discards a document's global structure and lets chunks compete with each other.
Instead of chunking, STAIR fine-tunes Mistral 7B to match a question against the document's full table of contents and name the section holding the answer — no vector database or embeddings needed, with the corpus stored in model parameters.
On their SearchTome benchmark (18 open textbooks, 6 fields, 103,249 test questions):
- 82.6% first-try section accuracy vs 76.9% for the same system without ToC (statistically significant in all 6 fields), 59.5% for BM25, 68.7% for DPR embeddings, 13.8% for untrained Mistral
- Invalid section names dropped from 3.25% to 0.05% — roughly 65x fewer, though partly by design since output is restricted to real section names
Limitations: the model must be trained per corpus, tests only covered textbooks, and questions were AI-generated rather than from real users. Paper and dataset are on arXiv (2609.03874).
Related event: IBM's STAIR Retriever Uses Document Structure to Slash Hallucinations(3 posts)→
More from Research
- BioCompute opens its remotely controlled wet lab, pitching 'Lovable for biology' — vaibhavbetter · 2026-11-04
- Legora CTO hand-writes 100k-param model weights with Fable, one-shot beats training — dair_ai · 2026-11-04
- 20-40% of alignment data is noisy: a loss-reduction scoring framework to fix it — burkov · 2026-11-04
- Scientists image quantum entangled photons in real time, and the result looks like a taijitu — CurieuxExplorer · 2026-11-04
- kalomaze: crude diffusion autoencoders may lose low-frequency info at the latent stage — kalomaze · 2026-11-04
- Structured captions boost text-to-music alignment, new Music DSG metric beats CLAP — affige_yang · 2026-11-04