A controlled experiment confirms representational entanglement causes collateral damage in unlearning.

Interpretability researchers have assumed shared structure between knowledge domains makes unlearning harder; nobody had tested it. The authors trained six 254M-parameter models on English Wikipedia with graded disentanglement between biology and non-biology knowledge, then applied three unlearning methods to each. At fixed forgetting, the most disentangled models incurred roughly 4x lower retain cost under two methods and 1.3x under the third. The intervention changes only the model, not the data or the algorithm, so this is causal evidence rather than correlation.