Does active sampling prevent textual degradation in recursive training?

large language models, model collapse, recursive training loop, textual degradation, active inheritance, synthetic data, áreas::ciências exatas e da terra::ciência da computação