Preventing model collapse may need 13-28x more fresh data than curing it, study finds

Published estimates for the fresh-data fraction needed to avoid model collapse when training on self-generated output range from about 1% to 30%, and the author argues this spread reflects different underlying questions rather than conflicting answers. A widely quoted closed-form expression for per-generation loss turns out to be only a lower bound on unconditional loss probability and does not measure what a training loop actually experiences. Simulations found that for symbols a loop almost always keeps, the folk constant understates loss by more than an order of magnitude, while for usually absent symbols it overstates it. The author also reports that accumulating training data across multiple generations, as most real pipelines do, moves the stability boundary by two to three orders of magnitude relative to simple replacement at the same fresh-data rate.