Summing Up the Facts: Additive Mechanisms Behind Factual Recall in LLMs

How do transformer-based large language models (LLMs) store and retrieve knowledge? We focus on the most basic form of this task -- factual recall, where the model is tasked with explicitly surfacing stored facts in prompts of form `Fact: The Colosseum is in the country of'. We find that the mechanistic story behind factual recall is more complex than previously thought. It comprises several distinct, independent, and qualitatively different mechanisms that additively combine, constructively interfering on the correct attribute. We …

basic country cs.lg facts focus form knowledge language language models large language large language models llms prompts recall store transformer