Decentralized Provenance Layer for Foundation Models: A Framework for Quantifying and Penalizing Synthetic Data Contamination in Recursive LLM Training

Abstract

The rapid development of Large Language Models (LLMs) and other Foundation Models requires substantial amounts of high-quality, human-generated training data [1]. However, the global collection of original human-generated content, known as the 'internet corpus,' is experiencing a marked decline [1]. To train successive generations of models, increasing quantities of data are sourced from the internet, which often includes outputs generated by existing AI models [1]. "Model collapse" describes a progressive degradation in learning, initiated when models are trained on data produced by their predecessors, resulting in a diminished ability to capture the true underlying data distribution [1]. This degradation is exacerbated by the loss of long-tail information, which is rare, difficult to obtain, and often sensitive. Consequently, model outputs become increasingly similar, reducing diversity and quality [1]. This issue impairs performance and may render future AI systems unreliable and less effective [1]. Recent research by Shumailov et al. (2024) highlights the model collapse phenomenon resulting from continual training on synthetic data. Borji (2024) further examines this issue using distribution fitting and iterative sampling of generated data [2].