Review Model Collapse

Model collapse is degradation that can occur when models are repeatedly trained on generated data in ways that cause the learned distribution to lose important variation, accuracy, or representation of rare cases.

If generated material replaces too much direct or well-curated evidence, later models may learn a narrower and increasingly distorted version of the original patterns.

Synthetic data can be useful, but its origin, quality, proportions, and effects need controls. Repeated reuse can hide what information has been lost.

Successive image models are trained largely on outputs from earlier models and gradually reproduce a narrower range of features while rare patterns disappear.

Model collapse is not an inevitable result of any synthetic data. It describes risks from particular recursive and poorly controlled training conditions.

Copies of copies can lose the original variation.

* [[synthetic-data|Synthetic Data]] * [[training-data|Training Data]] * [[data-provenance|Data Provenance]] * [[model-drift|Model Drift]]

* [Shumailov et al., AI models collapse when trained on recursively generated data](https://doi.org/10.1038/s41586-024-07566-y) — primary empirical paper on model collapse from recursive synthetic-data training.

* Track generated inputs with [[data-provenance|Data Provenance]] and distinguish this from [[model-drift|Model Drift]].