IBM's STAIR cuts hallucination 65x by having the LLM read the table of contents
alex_verem · x · 2026-10-14
IBM's STAIR (STructure Aware Information Retriever) ditches vector databases, embeddings, and re-rankers:
- Method: the model reads the question alongside the full table of contents and names the single section containing the answer—like a human using a book's ToC instead of cutting text into semantically arbitrary chunks
- Results: on 18 open textbooks across 6 fields, STAIR hits the right section 82.6% of the time vs 59.5% for keyword search and 45.6% for Claude Haiku 4.5; the strongest fine-tuned rival reached 76.9%, a statistically significant 5.7-point gap in all 6 fields
- Hallucination rate: only 0.05% invalid-section answers vs 3.25% for the rival—a 65x reduction
- Limits: requires documents with a ToC (textbooks, manuals, legal codes, enterprise docs; not email threads); built on Mistral 7B, runs on a single consumer GPU
Related event: IBM's STAIR uses document TOCs to cut hallucination rate to 0.05%(2 posts)→
More from Infra
- AMAT investor QA: inference-specific accelerators may not need 12-hi HBM stacks — BenBajarin · 2026-10-14
- Liquid AI hits 21k+ output tokens/s on NVIDIA B300 at high concurrency — JosephJacks_ · 2026-10-14
- Disaggregated Quantization lifts 1-bit Qwen3.8-27B from 29% to 61.5% on MMLU-Pro — udmrzn · 2026-10-14
- HINT4 targets a hardware-invariant NVFP4: 4-bit inference via int8 tensor cores on Ampere — secemp9 · 2026-10-14
- VigenFlow runs FLUX and Z-Image locally on AMD Ryzen AI NPU, no GPU needed — Kelsey_test · 2026-10-14
- RTX 5090 owner seeks free ComfyUI motion transfer workflows to avoid paid Higgsfield — Quirky_Fig3927 · 2026-10-14