Recall as the primary bottleneck for parametric factuality

Parametric factuality in generative models is currently limited by recall capabilities rather than reasoning. This suggests that improving model accuracy requires addressing how information is retrieved from internal weights.