aiqbench analysis
ASRU introduces a controllable unlearning framework for multimodal models that prioritizes generation quality alongside knowledge removal, addressing the common pitfall of producing hallucinated or rigid responses after unlearning. By combining activation steering with reinforcement learning, the method achieves a significant trade-off between forgetting sensitive data and preserving model utility, validated on Qwen3-VL with substantial gains in both metrics. This approach offers a practical solution for maintaining the safety and usability of MLLMs without requiring extensive retraining or supervision data.
ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
arXiv:2605.15687v2 Announce Type: replace Abstract: 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.
Abstract: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.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2605.15687 [cs.CL]
(or arXiv:2605.15687v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.15687 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 235 words
Related coverage
ACTIVE-o3: Empowering MLLMs with Active Perception via Pure Reinforcement Learning
arXiv · Jun 9, 2026
Active Inference as a Convex Markov Decision Process
arXiv · Jul 23, 2026
CAMeR: Keyword-Gated Hybrid Activation for Adaptive Memory Retention in LLM Agents
arXiv · Jul 25, 2026
Causal DAG-Guided Reinforcement Learning for Long-Term Physical Activity Maintenance: Using Prior Data to Inform Future Deployment
arXiv · Oct 28, 2026