Understanding Verbatim Memorization in LLMs Through Circuit Discovery

TL;DR

This paper investigates how large language models memorize data by analyzing transformer circuits, identifying specific mechanisms responsible for initiating and maintaining memorization, and examining how these processes transfer across contexts.

Contribution

It introduces a circuit-based interpretability approach to understand memorization in LLMs, revealing distinct circuits for initiation and maintenance of memorization.

Findings

Circuits that initiate memorization can also sustain it.

Memorization prevention mechanisms transfer across domains.

Memorization induction is highly context-dependent.

Abstract

Underlying mechanisms of memorization in LLMs -- the verbatim reproduction of training data -- remain poorly understood. What exact part of the network decides to retrieve a token that we would consider as start of memorization sequence? How exactly is the models' behaviour different when producing memorized sentence vs non-memorized? In this work we approach these questions from mechanistic interpretability standpoint by utilizing transformer circuits -- the minimal computational subgraphs that perform specific functions within the model. Through carefully constructed contrastive datasets, we identify points where model generation diverges from memorized content and isolate the specific circuits responsible for two distinct aspects of memorization. We find that circuits that initiate memorization can also maintain it once started, while circuits that only maintain memorization cannot trigger its initiation. Intriguingly, memorization prevention mechanisms transfer robustly across different text domains, while memorization induction appears more context-dependent.