CATA: Continual Machine Unlearning via Conflict-Averse Task Arithmetic

In three linesCATA introduces a continual machine unlearning method for vision-language models (VLMs). It represents each unlearning request as a task vector and aggregates historical vectors by suppressing conflicting components, ensuring forgetting effectiveness, model fidelity, and persistence against knowledge re-emergence.

Summary generated by Claude — human-verified