GoKawiilGoogle and Technion find frontier LLMs often 'know' facts they fail to recall
A new study from Google Research and Technion shows that when models like GPT-5 and Gemini-3 hallucinate, the underlying fact is usually already stored in their parameters rather than missing entirely. The researchers found these models encode 95-98% of tested facts, but often fail to surface them during generation due to retrieval, not storage, limits. They propose 'fact-level profiling' to separate whether a fact is encoded versus reliably retrievable across different phrasings.
GoKawiil's interpretation of the reporting above, not reported fact.
This challenges the common assumption that hallucinations mean models need more parameters, training data, or retrieval systems bolted on. If the bottleneck is recall rather than knowledge, then inference-time techniques like longer reasoning chains could meaningfully cut errors without costly retraining or scaling. It also suggests current accuracy benchmarks obscure the real cause of failures, making diagnosis and fixes harder for engineering teams.
- Frontier models encode 95-98% of tested facts but still fail to retrieve some correctly.
- Reasoning longer at inference time can recover up to 65% of facts models couldn't directly recall.
- Fact-level profiling distinguishes 'encoding' failures from 'recall' failures, pointing to different fixes.
Source: venturebeat.com, 2026-09-01
Published there as: “Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.