Published March 30, 2026
| Version v1.0
Preprint
Open
Model Collapse as Drift Cascade: Recursive AI Training, Entropy Loss, and the Thermodynamics of Synthetic Data Feedback
Description
Model collapse (Shumailov et al. Nature 2024) is the drift cascade operating on training data. Entropy loss 0.2-0.4/gen = drift velocity, data accumulation = constraint specification, first-order phase transition maps onto D1→D2→D3. 7/7 KC PASS.