Curated Synthetic Data Doesn't Have to Collapse: A Theoretical Study of Generative Retraining with Pluralistic Preferences
A new theoretical study published on arXiv explores how generative models can avoid collapse during recursive retraining. Researchers propose that using multiple, diverse reward functions for curation, rather than a single fixed one, can maintain output diversity. The study formalizes these dynamics, proving that under specific conditions, the model can converge to a stable distribution that balances competing preferences, akin to a Nash bargaining solution. AI
IMPACT Offers a theoretical framework to improve the stability and diversity of generative models during retraining.