Google study: hallucinations are lost keys, not empty shelves — CoT recovers up to 65% of facts
bendee983 · x · 2026-09-03
A new Google study suggests LLM hallucinations often stem not from missing knowledge ("empty shelves") but from retrieval failure ("lost keys") — the model can't retrieve facts it already knows.
The study profiles knowledge in five buckets, from facts retrievable directly to those needing special techniques, analogous to the human tip-of-the-tongue phenomenon. Experiments show frontier models can recover up to 65% of facts they fail to recall via direct inference simply by adding reasoning steps (CoT).
Practical takeaway: before reaching for RAG, try reasoning-based techniques to improve fact recall.
More from coding & agent
- Cohere Labs releases ATE dataset with ~700K tools from public MCP servers — Cohere_Labs · 2026-09-03
- Anthropic open-sources Claude Commerce Agents with retail and travel reference implementations — alex_verem · 2026-09-03
- Cohere Labs releases ATE dataset of ~700K tools from public MCP servers — Cohere_Labs · 2026-09-03
- Cursor's harness beats VS Code, but hooking it to local Ollama models needs workarounds — walkingriver · 2026-09-03
- Baseten ships remote MCP server and skill, cutting coding agents' token usage by up to 57% — baseten · 2026-09-03
- davidad: rollout rubrics in verifiable domains should check whether the model ran the verifier — davidad · 2026-09-03