Many-Shot Regurgitation (MSR) Prompting

TL;DR

This paper introduces Many-Shot Regurgitation (MSR) prompting, a black-box attack method to detect verbatim content reproduction in large language models by analyzing output patterns across different datasets.

Contribution

The paper presents MSR prompting as a novel framework for membership inference, revealing significant verbatim reproduction differences based on training data exposure in LLMs.

Findings

Higher verbatim match frequency in models' training data

Significant statistical differences between pre- and post-training datasets

Evidence of increased content reproduction when prompted with familiar data

Abstract

We introduce Many-Shot Regurgitation (MSR) prompting, a new black-box membership inference attack framework for examining verbatim content reproduction in large language models (LLMs). MSR prompting involves dividing the input text into multiple segments and creating a single prompt that includes a series of faux conversation rounds between a user and a language model to elicit verbatim regurgitation. We apply MSR prompting to diverse text sources, including Wikipedia articles and open educational resources (OER) textbooks, which provide high-quality, factual content and are continuously updated over time. For each source, we curate two dataset types: one that LLMs were likely exposed to during training ($D_{\rm pre}$) and another consisting of documents published after the models' training cutoff dates ($D_{\rm post}$). To quantify the occurrence of verbatim matches, we employ the Longest Common Substring algorithm and count the frequency of matches at different length thresholds. We then use statistical measures such as Cliff's delta, Kolmogorov-Smirnov (KS) distance, and Kruskal-Wallis H test to determine whether the distribution of verbatim matches differs significantly between $D_{\rm pre}$ and $D_{\rm post}$. Our findings reveal a striking difference in the distribution of verbatim matches between $D_{\rm pre}$ and $D_{\rm post}$, with the frequency of verbatim reproduction being significantly higher when LLMs (e.g. GPT models and LLaMAs) are prompted with text from datasets they were likely trained on. For instance, when using GPT-3.5 on Wikipedia articles, we observe a substantial effect size (Cliff's delta $= -0.984$) and a large KS distance ($0.875$) between the distributions of $D_{\rm pre}$ and $D_{\rm post}$. Our results provide compelling evidence that LLMs are more prone to reproducing verbatim content when the input text is likely sourced from their training data.