"Recall Is the Bottleneck for Parametric Factuality" (v2 June 19) with its companion dataset WikiProfile: 2,150 Wikipedia facts, each probed under 10 phrasings (21,500 instances, CC-BY-SA-4.0). Core finding: models encode 95–98% of tested facts somewhere in their parameters, but recall — retrieving the fact under a given phrasing — is the bottleneck for factuality, reframing how hallucination should be attacked.

Paper

factualityresearchdataset