Bilayer SIR model explains AI model collapse from synthetic data

Arxiv CS.CL (NLP/language models)

A new arXiv paper introduces a bilayer SIR model to study cross-contamination in AI training with synthetic data. The model shows that when models train on data from other models, collapse occurs faster than single-chain degradation. This provides a framework for understanding ecosystem-level risks.

model-collapsesynthetic-data

Arxiv CS.CL (NLP/language models)·Xiangyu Wang

6 days ago