From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources

A new statistical study challenges the prevailing fear of irreversible 'model collapse' in AI. The research argues that iterative training on data contaminated with synthetic outputs can actually lead to improvement, not degradation, as long as a non-trivial amount of fresh, real data is included in the mix. The analysis shows that with the right balance of mixture weights and sample sizes, models can avoid collapse and even recover the true target distribution. Simulation studies support these findings across different model classes.