When a model is trained on data and later asked to forget some of it, the usual approach is to measure whether the updated model's predictions match a model retrained from scratch on the remaining data. This snapshot check can hide important differences in the internal features the two models learn. Two models that make the same predictions may still rely on different patterns in the data, and those differences can become visible when the model is adapted or trained further.
The paper proves that if an unlearned model and a retrained model encode different features, a single update to the output layer can break their agreement. It also shows that deleting even a tiny fraction of data can cause retraining to learn a new feature the original model never had. The researchers introduce a way to measure how well an unlearned model recovers the features of the retrained model, separately tracking missing features and extra ones.
Testing eight common unlearning methods, they found that many that perform well on standard tests still fail to recover the retrained features. This mismatch predicts how far the model will deviate from retraining when it is later trained on new data or adapted to a new task. Directly fixing the feature mismatch reduces these deviations, suggesting that unlearning should be evaluated by feature recovery, not just immediate prediction agreement.