Skip to content
BENIBIX Learn

Model Collapse

AI💬 LLMsLanguage Model Training

Also known as AI Model Collapse, Recursive Training Collapse

Models trained on other models' text get worse each round.

Think of it like

It is like a copy of a copy of a copy. Each one is a bit more blurry.

Example

Researchers train each new model only on text from the one before. After a few rounds, the rare words and odd facts are gone, and all the text sounds alike.

Why it matters

More and more of the web is written by AI. Future models trained on it could take in this slow decay.

Definition

Model collapse is the harm that builds up when models learn, round after round, from text made by earlier models. Each model copies the common patterns well but drops some rare ones. Over many rounds, the rare parts of the first data fade away. The output grows narrow and dull. A 2024 paper in Nature showed it in language models and in other kinds of model. Keeping real human data in the mix, and filtering made-up data, lowers the risk.
Each round trains on the last model's output. The rare edges shrink until little is left.

Common mistake

Thinking all made-up training data is harmful. Checked and mixed with real data, it is widely used.

Key takeaways

  1. 1
    Copies of copies. Each round loses a little.
  2. 2
    Rare things go first. The edges of the data fade.
  3. 3
    Keep real data. Fresh human data slows the decay.

Compare

Learn it properly