Answers are grounded in this paper and cite their sources.
Demonstrates forgetting certificates do not survive model merging, indicating a need for new methods.
Machine unlearning sells a guarantee: hand over the forbidden data and the model will be made to behave as though it never saw it — forgotten for you, wholesale. The guarantee is written on the unlearned model alone: forget-set accuracy at chance, a membership-inference null, a probe that no longer reads the concept. We show the guarantee is void the instant the model is merged. Weight-space merging (model soups, task arithmetic, TIES, DARE) is now standard, and it composes weights, not behaviours. We give the exact law of the failure: after a merge, the signal along the forget direction equals the merge coefficient times the partner's own component along that direction, and is completely independent of how thoroughly the first model forgot. Three things follow, each proved and each confirmed on trained networks with planted ground truth. (i) An impossibility: no certificate that reads only the unlearned model can survive a merge — certified forgetting is not compositional. (ii) A tight boundary: the merge is safe if and only if the partner is orthogonal to the forget direction — exact, checkable, and a fact about the partner rather than about the model you certified. (iii) A fix: unlearn in the shared base, not in the fine-tuning delta, and forgetting survives every orthogonal merge. On a planted-fact benchmark across five seeds, delta-space unlearning reaches chance with utility intact; a routine merge with a knowledgeable partner then drags the fact back.
No takes yet. Share an insight, caveat, or question.
Madison Charlton (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: