Is Forgetting Unified? Exploring Concept Unlearning in Unified Multimodal Models

Is Forgetting Unified? Exploring Concept Unlearning in Unified Multimodal Models. By Zhongkai Xue and nine others, from the Institute of Automation, Chinese Academy of Sciences, the University of Chinese Academy of Sciences, and three other institutions. Published on the third of November.

Unified multimodal models do two jobs in one: they generate images, and they understand images they are shown. Unlearning methods, though, have mostly been built for one pathway at a time. The question is whether a single-pathway method is enough to suppress a target concept on both sides at once.

The goal studied is cross-pathway concept unlearning. A model should neither generate images of the target concept nor identify that concept in input images, while keeping its other capabilities intact. What gets measured is joint forgetting across both pathways, alongside how much capability is preserved.

The benchmark is UUBench, a diagnostic set covering thirty target concepts and sixty semantically related neighboring concepts. The neighbors are there because forgetting a concept should not drag nearby concepts down with it.

On representative unified models, the evaluated single-pathway methods achieve only limited joint forgetting. Editors that work on the generation side strongly suppress generating the target, but forgetting in visual understanding stays limited. Baselines that work on the understanding side struggle to forget much even within their own pathway. Simply combining the two kinds of method also gives limited joint forgetting. The authors then propose UniForget, which reaches joint forgetting rates of eighty percent on BAGEL-7B-MoT and ninety-two point seven percent on Janus-Pro-7B, ahead of representative baselines. Further analysis shows trade-offs between joint forgetting and capability preservation.

The title asks whether forgetting is unified, and on this evidence it is not, at least not for free. Removing a concept from what a model draws does not remove it from what the model recognizes, and closing that gap comes with a cost to the capabilities the model is supposed to keep.