26 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Exact Unlearning of Finetuning Data via Model Merging at Scale
Carnegie Mellon University
Recording
Abstract
Kevin Kuo presents SIFT-Masks, a model-merging approach for exactly removing contributions from fine-tuned models. Local masks preserve task-specific utility, while a common sign constraint lets the masks be computed independently. Experiments merging up to 500 models compare accuracy with naive merging and measure the computational cost of exact removal against other merging methods.