When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting
The paper argues that some finetuning “forgetting” is a reversible access failure, not erased knowledge.
Old facts can disappear during training on new facts, then return even without retraining on the old data. The authors model this as a shared shift in old representations: their geometry stays intact, but recall is hidden. In their tests, removing that common direction prevents the collapse in a synthetic Transformer and restores old facts in a pretrained model. The lasting damage comes from slower, fact-specific erosion. ArXiv · AI/CL/LG's note
Old facts can disappear during training on new facts, then return even without retraining on the old data. The authors model this as a shared shift in old representations: their geometry stays intact, but recall is hidden. In their tests, removing that common direction prevents the collapse in a synthetic Transformer and restores old facts in a pretrained model. The lasting damage comes from slower, fact-specific erosion. ArXiv · AI/CL/LG's note
score 5