Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality

Standard factuality evaluations of LLMs treat all errors alike, obscuring whether failures arise from missing knowledge (empty shelves) or from limited access to encoded facts (lost keys). We propose a behavioral framework that profiles factual knowledge at the level of facts rather than questions, characterizing each fact by whether it is encoded, and then by how accessible it is: cannot be recalled, can be directly recalled, or can only be recalled with inference-time computation (thinking). To support such profiling, we introduce WikiProfile, a new benchmark constructed via an automated pipeline with a prompted LLM grounded in web search. Across 4 million responses from 13 LLMs, we find that encoding is nearly saturated in frontier models on our benchmark, with GPT-5 and Gemini-3 encoding 95--98\% of facts. However, recall remains a major bottleneck: many errors previously attributed to missing knowledge instead stem from failures to access it. These failures are systematic and disproportionately affect long-tail facts and reverse questions. Finally, we show that thinking improves recall and can recover a substantial fraction of failures, indicating that future gains may rely less on scaling and more on methods that improve how models utilize what they already encode.

When LLMs get a fact wrong, we usually just mark it as an error and move on. But not all errors are the same: sometimes the model never learned the fact in the first place (i.e., it does not encode the fact), and sometimes it did learn it but simply can't recall it, like knowing a word but having it stuck on the tip of your tongue. Current evaluations can't tell these two failure types apart. We developed a framework that classifies each fact a model knows (or doesn't) into one of five profiles, distinguishing between facts that were never learned, facts that are learned but cannot be recalled, and facts that can be. We evaluated 13 leading LLMs using a new benchmark of 2,150 facts. Our results show that frontier LLMs have encoded almost every fact, 95–98% of the tested facts. The real bottleneck is recall. Crucially, when LLMs "think through" a question, they can recover up to 65% of these recall failures. This suggests that future improvements in factuality should focus less on training larger models and more on teaching them better access to what they already know.