A mistake where a model repeats training examples instead of making a fresh answer.
Model memorization is like a kid who studied by photocopying the answer key. Give it the first line, and it blurts out the whole page, coffee stain included.
This can leak private data or copyrighted text. It can also make test scores look fake if the model saw the questions before.
Data-privacy
Model memorization can reveal private details from training data.
Differential Privacy
Differential Privacy adds noise during training, so memorization is less likely.
Overparameterization
Overparameterization gives the model more room to memorize examples.
Benchmark contamination
Benchmark contamination happens when memorized test questions make scores look too high.