Recall, Not Encoding, Limits Fact Accuracy in Gemini3 and GPT‑5
The team built WikiProfile – 2,150 Wikipedia facts each paired with ten probing questions – to separate fact encoding from recall. Their analysis shows that Gemini3 and GPT‑5 store most facts but frequently fail to retrieve them, making recall the primary bottleneck for parametric factuality. Post‑training retrieval tricks could boost factual accuracy more than raw scaling.