When Unlearning Fails: Reliable Data Deletion under Post-Training in Agent Networks
MUTE proposes a method to reliably delete data from self-improving federated agent networks by tracing downstream influence and removing model and data residue. It works by estimating how forgotten data influences future behavior, erasing current model residue, quarantining high-influence retained trajectories, and auditing regeneration under an uplink budget to ensure deletion persists.