Routing-Aware Expert Calibration for Machine Unlearning in Mixture-of-Experts Language Models

In three linesTRACE, an unlearning method for MoE models, detects forget-critical experts activated by data to unlearn and recalibrates retention regularization by reweighting token-level losses to match expert activation frequencies. Experiments on WMDP and MUSE-BOOKS show 9% relative utility improvement over strongest baseline.

Summary generated by Claude — human-verified