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

TL;DR

This paper investigates how large language models recall factual information, revealing that multiple independent mechanisms additively contribute to correct answers, and introduces methods to analyze these mechanisms and attention heads.

Contribution

It uncovers the additive mechanisms behind factual recall in LLMs and extends attribution techniques to better understand attention head contributions.

Findings

Factual recall involves multiple independent additive mechanisms.

Mechanisms interfere constructively to produce correct answers.

Extended attribution methods reveal mixed attention heads from different source tokens.

Abstract

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 term this generic phenomena the additive motif: models compute through summing up multiple independent contributions. Each mechanism's contribution may be insufficient alone, but summing results in constructive interfere on the correct answer. In addition, we extend the method of direct logit attribution to attribute an attention head's output to individual source tokens. We use this technique to unpack what we call `mixed heads' -- which are themselves a pair of two separate additive updates from different source tokens.