Saved in:
Bibliographic Details
Main Authors: Guang, Jiahui, Zhu, Yingjie, Gao, Cuiyun, Wang, Haiyan, Li, Jing, Shao, Di, Gu, Zhaoquan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.15687
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910223542779904
author Guang, Jiahui
Zhu, Yingjie
Gao, Cuiyun
Wang, Haiyan
Li, Jing
Shao, Di
Gu, Zhaoquan
author_facet Guang, Jiahui
Zhu, Yingjie
Gao, Cuiyun
Wang, Haiyan
Li, Jing
Shao, Di
Gu, Zhaoquan
contents Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.
format Preprint
id arxiv_https___arxiv_org_abs_2605_15687
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
Guang, Jiahui
Zhu, Yingjie
Gao, Cuiyun
Wang, Haiyan
Li, Jing
Shao, Di
Gu, Zhaoquan
Computation and Language
Artificial Intelligence
Multimodal large language models (MLLMs) may memorize sensitive cross-modal information during pretraining, making machine unlearning (MU) crucial. Existing methods typically evaluate unlearning effectiveness based on output deviations, while overlooking the generation quality after unlearning. This can easily lead to hallucinated or rigid responses, thereby affecting the usability and safety of the unlearned model. To address this issue, we propose ASRU, a controllable multimodal unlearning framework that incorporates generation quality as a core evaluation objective. ASRU first induces initial refusal behavior through activation redirection, and then optimizes fine-grained refusal boundaries using a customized reward function, thereby achieving a better trade-off between target knowledge unlearning and model utility. Experiments on Qwen3-VL show that ASRU significantly improves unlearning effectiveness (+24.6%) on average and generation quality (5.8x) on average while effectively preserving model utility, using only a small amount of retained supervision data.
title ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2605.15687