40pct-probe-vs-generation-gap

IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-s4-intra-memory-conflict.md

Created 2026-08-25T02:59:03+00:00

Knowledge-probe accuracy exceeds generation accuracy by approximately 40%, indicating factual knowledge is stored in parameters but not reliably retrieved during decoding (Li et al., 2023c).

Summary

The model actually knows more than it shows: it scores roughly 40% higher on direct recall tests than on free-form answers, meaning the facts are sitting in its weights but something in the generation process lets them slip. The practical implication is that closing the output gap — getting the model to surface what it already stores — should buy more accuracy than feeding it new information.