Definition
Plain language
A test of whether a model has truly forgotten something — you give it a tiny bit of retraining and see if the deleted content comes flooding back.
As stated in the literature
An attack on machine-unlearning methods that briefly fine-tunes the unlearned model and measures how fast suppressed content recovers; post-hoc unlearning typically fails within a few gradient steps, while natively-unlearnable architectures track a never-trained baseline.
Also called: relearning attacks
Why it matters: It tests whether deleted knowledge is truly gone or merely hidden, exposing unlearning methods that only paper over content that can quickly resurface.
For example, you give a supposedly 'forgotten' fact just a few rounds of retraining and watch whether the model suddenly remembers it again.
Heard on the show
“First attack: the relearning attack, the same one that broke post-hoc unlearning in under ten steps.”Episode 145 — Building Forgetting Into a Language Model With One Extra Line of Code