ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models

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 (5.8×) on average while effectively preserving model utility, using only a small amount of retained supervision data.

Multimodal large language models (MLLMs) can answer questions about both images and text, but they may also memorize sensitive information during training, such as private details linked to a person’s image. When such information needs to be removed, simply making the model “forget” is not enough: the model should also respond naturally and safely instead of hallucinating facts or giving rigid, repetitive refusals. We propose ASRU, a method that helps multimodal models forget target information while still remaining useful. ASRU first guides the model toward a basic refusal behavior, and then further trains it to decide more carefully when to refuse and when to answer. This allows the model to give context-appropriate refusals for forgotten information while preserving its ability to answer normal questions. Our experiments show that ASRU improves both unlearning effectiveness and the quality of responses after unlearning, while using only a small amount of retained supervision data. This work helps make MLLMs more privacy-aware, reliable, and practical after sensitive information has been removed.