Unlearning

Unlearning refers to techniques that attempt to remove specific knowledge or capabilities from a trained model without retraining it from scratch, typically by fine-tuning against the targeted information (as in methods like RMU or NPO). It serves as a key test case for interpretability tools: if a model has genuinely forgotten something, a probe looking for that knowledge in its internals should fail to find it, whereas residual traces would suggest the information was merely suppressed rather than erased.