1 paper
Ilya Lasy, Peter Knees, Stefan Woltran
Underlying mechanisms of memorization in LLMs -- the verbatim reproduction of training data -- remain poorly understood. What exact part of the network decides to retrieve a token…