Is Forgetting Unified? Exploring Concept Unlearning in Unified Multimodal Models
AuthorsZhongkai Xue, Junkai Chen, Mingcong Lei, Rumo Zhang, Yixiao Liu, Zhimin Sun, Shu Wu, Qiang Liu, Lihua Zhang, Liang Wang
AffiliationsInstitute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · ByteDance · Shanghai Jiao Tong University · The Chinese University of Hong Kong, Shenzhen
UniForget aims to make unified multimodal models both stop generating and stop recognizing selected concepts while preserving other capabilities.
Key results
UNIFORGET joint forgetting rate
UNIFORGET joint forgetting rate
Rate requiring coherent generation and useful visual descriptions
Rate requiring coherent generation and useful visual descriptions
Images across target and neighboring concepts
What the paper found
This paper asks whether deleting a visual concept from a unified multimodal model also stops it from recognizing that concept. Its UUBENCH benchmark pairs image generation with visual question answering across 30 target concepts, 60 neighboring concepts, and 720 reference images. The proposed UNIFORGET method combines closed-form generation-side weight editing with LoRA updates for understanding, training the model to avoid naming a target while still describing its visible attributes. On BAGEL-7B-MoT and Janus-Pro-7B, UNIFORGET reaches 80.0% and 92.7% Unified Forget, compared with 31.3% and 41.3% for the direct TIME plus gradient-descent baseline. When the metric also requires coherent generation and useful visual descriptions, the rates drop to 47.6% and 58.0%, exposing a trade-off between forgetting and preserving capability. The results are relevant to companies such as ByteDance working with unified multimodal AI; evaluation used OpenAI’s gpt-5.4-mini as an automated judge. The study’s central finding is that suppressing image generation alone does not reliably erase visual recognition, so both pathways need coordinated evaluation and editing.
Original abstract
Unified multimodal models (UMMs) integrate image generation and visual understanding, but are existing single-pathway unlearning methods sufficient to suppress target concepts across both? In this work, we study cross-pathway concept unlearning. Our goal is to prevent a model from both generating images of a target concept and identifying it in input images, while preserving other capabilities. To this end, we introduce UUBench, a diagnostic benchmark covering 30 target concepts and 60 semantically related neighboring concepts. Experiments on representative UMMs show that the evaluated single-pathway methods achieve limited joint forgetting: generation-side editors strongly suppress target generation but yield limited forgetting in visual understanding, while understanding-side baselines struggle to achieve substantial forgetting even in their own pathway. Their direct composition also yields limited joint forgetting. We therefore propose UniForget, achieving joint forgetting rates of 80.0% on BAGEL-7B-MoT and 92.7% on Janus-Pro-7B, outperforming representative baselines. Further analysis reveals trade-offs between joint forgetting and capability preservation.
Read the original paperMore in AI Safety
Browse all 54 papers →LEASH: Stay Safe in Agentic Robotics
Qinfeng Zhu, Weiguang Zhao, Ke Sun, Xihao Guo, Yunxi Jiang, Rui Zhang, Lei Fan, Kaizhu Huang
RoboRisk tests how safely AI agents control robots, while LEASH adds checks intended to prevent dangerous plans and actions.
"Shall We Approve?" Consequence-Aware Approval for Autonomous Coding Agents
Weiheng Wu, Junkai Chen, Zhensu Sun, Jieke Shi, Yunbo Lyu, David Lo
Caper tests what a coding agent’s command actually does in a sandbox before approving it, aiming to block harmful actions without stopping safe ones.
How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates
Max Neuwinger, Hanna Yukhymenko, Imanol Schlag
This paper uses culturally informed templates to expand safety training and testing for language models across 60 languages.