Statistical Knowledge Assessment for Generative Language Models

Generative Language Models (GLMs) have demonstrated capabilities to store factual knowledge and answer queries efficiently. Given varying prompts, does a GLM consistently generate factually correct answers? In this paper, we introduce a statistical knowledge assessment framework guided by latent variables and the KaRR metric, which quantifies a model's knowledge by computing its continuous probability across diverse text forms. We conduct a comprehensive comparison of knowledge across 14 GLMs using our framework, including LLaMA, Alpaca, OPT, and others. Our statistical knowledge assessment encompasses 600 relation types and exhibits a strong correlation (0.43 Kendall's $\tau$) with human evaluation. Our findings reveal that the knowledge in GLMs with the same backbone architecture adheres to the scaling law, and that tuning on instruction-following data may compromise the model's ability to generate factually correct text consistently.