ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
For practitioners deploying MLLMs, this method addresses the overlooked issue of generation quality degradation after unlearning, improving both safety and usability.
ASRU introduces a controllable unlearning framework for MLLMs that balances target knowledge removal with generation quality, achieving a 24.6% improvement in unlearning effectiveness and 5.8x better generation quality on Qwen3-VL.
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.