Forgetting isn't unified in unified multimodal models: editing only the image path leaves 24% recognition forgetting, while UniForget reaches 80% joint forgetting
Synopsis
Unified multimodal models generate images and understand them in one model, yet existing single-pathway unlearning methods struggle to make a target concept disappear from both pathways. UUBench pairs generation and understanding tasks around 30 target concepts and 60 same-category neighbors, and UniForget merges generation-side weight editing with understanding-side low-rank adapter updates into one checkpoint, reaching 80.0% joint forgetting on BAGEL-7B-MoT and 92.7% on Janus-Pro-7B while exposing a trade-off between forgetting and capability preservation.
Interpretation
A target concept has to be suppressed in both image generation and visual understanding, while existing methods mostly handle one pathway. Earlier concept-unlearning work concentrated on a single channel: either suppressing target generation in vision generative models or erasing sensitive factual knowledge from language and multimodal understanding models. On BAGEL-7B-MoT the generation-side method TIME reaches 100.0% target forgetting but only 24.0% direct forgetting in understanding; on Janus-Pro-7B the two figures are 96.7% and 30.7%.
UniForget first applies a closed-form weight edit to the generation pathway, then adjusts the understanding pathway with low-rank adapters over a frozen backbone, and finally merges both updates into one checkpoint. Directly combining the strongest generation-side method TIME with the understanding-side method GD yields only limited joint forgetting. TIME + GD reaches 31.3% joint forgetting on BAGEL-7B-MoT and 41.3% on Janus-Pro-7B, against 80.0% and 92.7% for UniForget, which also keeps 100.0% target forgetting on both backbones.
Joint forgetting and capability preservation show a measurable trade-off, and the two backbones behave differently. Under the strict criterion that also requires image quality and informative answers, joint forgetting is clearly lower than plain joint forgetting. UniForget's strict joint forgetting is 47.6% on BAGEL-7B-MoT and 58.0% on Janus-Pro-7B; strict neighbor retention is 84.6% on BAGEL and 73.2% on Janus.
Perspective
The evaluation is limited to two 7B unified multimodal models, BAGEL-7B-MoT and Janus-Pro-7B, and to the 30 target concepts and 60 neighboring concepts of UUBench, with image quality and visual utility scored by automated metrics and LLM judges.
The authors note that experiments cover only two 7B models and 30 target concepts, leaving generalization to larger models and broader domains to future work; automated metrics and LLM judges may differ from human judgment; and the evaluation targets behavioral forgetting under standard prompts, with certified deletion from model weights beyond its scope. When scaling to multiple concepts, understanding-side forgetting drops as the concept count grows, and name suppression under prompt attacks shows model-specific sensitivity.