{"id":"arxiv/2602.10531","url":"https://arxiv.org/abs/2602.10531","tags":["stat.ML","cs.LG","math.ST"],"score":0,"title":"From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources","domain":"stat.ML","source":"arxiv","authors":["Soham Bakshi","Sunrit Chakraborty"],"reviews":0,"subject":"stat.ML","summary":"The problem of model collapse has presented new challenges in iterative training of generative models, where such training with synthetic data leads to an overall degradation of performance. This paper looks at the problem from a statistical viewpoint, illustrating that one can actually hope for improvement when models are trained on data contaminated with synthetic samples, as long as there is some amount of fresh information from the true target distribution. In particular, we consider iterative training on samples sourced from a mixture of the true target and synthetic distributions. We…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-02-11","reactions":{}}