Papers

Memorization lives in a few tokens.

Mitigating Memorization Where It Happens · NeurIPS 2026

Language models sometimes recite their training data word for word. Most fixes retrain the model or slow down every token it writes. We find the tokens where recitation is happening and steer the model away there, and nowhere else. It matches or beats the best existing methods at 100–1000× less compute, without changing any weights.