Factuality is essential for reliable LLMs. Standard accuracy metrics collapse encoding failures (fact not stored) and recall failures (fact stored but inaccessible). The article introduces knowledge profiling, a behavioral framework to distinguish these, and applies it to frontier LLMs like Gemini-3-Pro and GPT-5.

Using the Wiki Profile benchmark of 2,150 Wikipedia-derived facts, the study finds that 95-98% of facts are encoded in frontier models, yet 26-34% cannot be directly recalled. Even with thinking, 11-12% still fail. This reveals that the primary bottleneck for parametric factuality is recall, not encoding.

Recall failures are tied to training context: rare facts are encoded but harder to recall, and the reversal curse (knowing "A is B" but not reverse) is a recall problem, as models recognize the answer in multiple-choice. Thinking (chain-of-thought) improves recall most where direct recall is weakest, offering a recovery mechanism.