ROME: Memorization Insights from Text, Logits and Representation

TL;DR

This paper introduces ROME, a novel method for analyzing memorization in language models by examining logits and representations without needing extensive training data processing, revealing key factors influencing memorization.

Contribution

The paper proposes ROME, an innovative approach that analyzes memorization through model outputs and internal representations, bypassing the need for direct training data examination.

Findings

Longer words are less likely to be memorized

Higher confidence correlates with greater memorization

Representations of the same concepts are more similar across contexts

Abstract

Previous works have evaluated memorization by comparing model outputs with training corpora, examining how factors such as data duplication, model size, and prompt length influence memorization. However, analyzing these extensive training corpora is highly time-consuming. To address this challenge, this paper proposes an innovative approach named ROME that bypasses direct processing of the training data. Specifically, we select datasets categorized into three distinct types -- context-independent, conventional, and factual -- and redefine memorization as the ability to produce correct answers under these conditions. Our analysis then focuses on disparities between memorized and non-memorized samples by examining the logits and representations of generated texts. Experimental findings reveal that longer words are less likely to be memorized, higher confidence correlates with greater memorization, and representations of the same concepts are more similar across different contexts. Our code and data will be publicly available when the paper is accepted.