| \documentclass[11pt,letterpaper]{article} |
|
|
| \usepackage[margin=1.2in]{geometry} |
| \usepackage[skip=6pt]{parskip} |
|
|
| \usepackage{microtype} |
| \usepackage{graphicx} |
| \usepackage{subcaption} |
| \usepackage{booktabs} |
| \usepackage{xcolor} |
|
|
| \definecolor{paperblue}{rgb}{0,0.10,0.45} |
| \usepackage[ |
| colorlinks=true, |
| linkcolor=paperblue, |
| citecolor=paperblue, |
| urlcolor=paperblue |
| ]{hyperref} |
|
|
| \newcommand{\theHalgorithm}{\arabic{algorithm}} |
|
|
| \usepackage{amsmath} |
| \usepackage{amssymb} |
| \usepackage{mathtools} |
| \usepackage{amsthm} |
|
|
| \usepackage[capitalize,noabbrev]{cleveref} |
|
|
| \theoremstyle{plain} |
| \newtheorem{theorem}{Theorem}[section] |
| \newtheorem{proposition}[theorem]{Proposition} |
| \newtheorem{lemma}[theorem]{Lemma} |
| \newtheorem{corollary}[theorem]{Corollary} |
| \theoremstyle{definition} |
| \newtheorem{definition}[theorem]{Definition} |
| \newtheorem{assumption}[theorem]{Assumption} |
| \theoremstyle{remark} |
| \newtheorem{remark}[theorem]{Remark} |
|
|
| \Crefname{definition}{Definition}{Definitions} |
|
|
| \newtheorem{claim}{Claim} |
| \newtheorem{condition}{Condition} |
| \newtheorem*{definition*}{Definition} |
|
|
| \usepackage{thm-restate} |
| \usepackage{enumitem} |
| \usepackage{tabularx} |
| \usepackage{tikz} |
|
|
| \input{macros} |
|
|
| \RequirePackage{algorithm} |
| \RequirePackage{algpseudocodex} |
| \RequirePackage{natbib} |
| \RequirePackage{eso-pic} |
| \RequirePackage{forloop} |
| \RequirePackage{url} |
| \RequirePackage{caption} |
|
|
|
|
| \title{Language Generation with Replay:\\A Learning-Theoretic View of Model Collapse} |
| \author{ |
| Giorgio Racca\\ |
| University of Copenhagen\\ |
| \texttt{g.racca@di.ku.dk} |
| \and |
| Michal Valko\\ |
| Isara Labs\\ |
| \texttt{michal@isara.io} |
| \and |
| Amartya Sanyal\\ |
| University of Copenhagen\\ |
| \texttt{amsa@di.ku.dk} |
| } |
| \date{} |
|
|
|
|
| \begin{document} |
| \maketitle |
|
|
| \begin{abstract} |
| As scaling laws push the training of frontier large language models (LLMs) toward ever-growing data requirements, training pipelines are approaching a regime where much of the publicly available online text may be consumed. |
| At the same time, widespread LLM usage increases the volume of machine-generated content on the web; together, these trends raise the likelihood of generated text re-entering future training corpora, increasing the associated risk of performance degradation often called \emph{model collapse}. |
| In practice, model developers address this concern through data cleaning, watermarking, synthetic-data policies, or, in some cases, blissful ignorance. |
| However, the problem of model collapse in generative models has not been examined from a learning-theoretic perspective: we study it through the theoretical lens of the \emph{language generation in the limit} framework, introducing a \emph{replay} adversary that augments the example stream with the generator's own past outputs. |
| Our main contribution is a fine-grained learning-theoretic characterization of when replay fundamentally limits generation: while replay is benign for the strongest notion of uniform generation, it provably creates separations for the weaker notions of non-uniform generation and generation in the limit. |
| Interestingly, our positive results mirror heuristics widely used in practice, such as data cleaning, watermarking, and output filtering, while our separations show when these ideas can fail. |
| \end{abstract} |
|
|
| \input{sections/01-intro} |
| \input{sections/02-related-work.tex} |
| \input{sections/03-setup-results} |
| \input{sections/04-uniform-with-replay} |
| \input{sections/05-non-uniform-with-replay} |
| \input{sections/06-in-the-limit-with-replay} |
| \input{sections/07-proper} |
| \input{sections/08-discussion} |
|
|
| \section*{Acknowledgements} |
| GR and AS acknowledge the Novo Nordisk Foundation for support via the Startup grant (NNF\allowbreak24OC0087820); AS additionally acknowledges support from VILLUM FONDEN via the Young Investigator program (VIL72069). |
| The authors also thank Carolin Heinzler for helpful feedback, as well as the anonymous reviewers for their constructive comments. |
|
|
| \begingroup |
| \setlength{\bibsep}{5.85pt plus 0.3pt} |
| \bibliographystyle{plainnat} |
| \bibliography{ref} |
| \endgroup |
|
|
| \newpage |
| \appendix |
| \onecolumn |
|
|
| \input{appendix/A-overview-generation} |
|
|
| \end{document} |
|
|