Knowledgeless Language Models cut closed-book recall by anonymizing entities during pretraining

gdm3000 · x · 2026-07-21

The paper proposes Knowledgeless Language Models (KLLMs), a new pretraining paradigm designed to reduce parametric recall and push models toward evidence-grounded reasoning.

The key idea is to anonymize named entities during pretraining, removing a major channel for entity-linked factual supervision. The authors argue this substantially lowers closed-book factual recall, while often improving performance when relevant information is provided in context. The screenshot emphasizes the broader motivation: language models store substantial factual knowledge, but that can become unreliable when the knowledge is outdated, incomplete, or mismatched with the prompt.

Original post →

More from Research

Research channel →