Curated Synthetic Data Prevents Model Collapse in Generative Retraining

A recent theoretical investigation published on arXiv in May 2025 disputes the common belief that recursive retraining of generative models inevitably results in output collapse. The authors demonstrate that by curating synthetic outputs using various reward functions instead of relying on a single, fixed signal, the model can preserve its diversity. This research articulates the dynamics of recursive training amidst differing preferences and establishes convergence towards a stable distribution that spreads probability across competing high-reward areas. This limiting distribution aligns with a weighted Nash bargaining solution, giving a formal interpretation of value aggregation within synthetic retraining processes. The findings challenge earlier assertions that collapse is unavoidable without incorporating real data, thus supporting the use of curated synthetic data in AI training frameworks.