Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ASRU, a two-stage multimodal unlearning method that first steers activations toward refusal by updating one MLP down-projection, then refines the forget-retain refusal boundary with GRPO and a rule-based reward, evaluated on MLLMU-Bench and CLEAR with Qwen3-VL and LLaVA.

Problem

MLLM unlearning methods are evaluated mainly by output deviation on the forget set, ignoring post-unlearning generation quality, which yields hallucinated, rigid, or degenerate responses and a poor forget-utility trade-off.

Approach

Combine activation steering (a closed-form single down-projection update toward a knowledge-absence refusal direction) with GRPO on forget samples paired with a small set of similar retain "boundary" samples, using verifiable rule-based rewards that reward refusal on forget queries and correct answers on boundary queries.

Key results

• On Qwen3-VL-8B (5% forget), ASRU reduces forget classification accuracy to 30.83 vs 43.20 for the best baseline MMUnlearner and retains 53.74 retain accuracy vs 50.66 for KL_Min.
• ASRU attains a contextual-refusal score of 2.98 with GPT-4o-mini versus at most 0.68 for baselines on Qwen3-VL-8B at 5% forget.
• Ablation: removing the boundary set drops retain accuracy to 28.95, showing over-forgetting and utility collapse.

Surprising findings

• Removing the boundary set yields the strongest refusals (higher generation-quality score of 4.18) but severe utility collapse.
• Mixed-context training (ASRU_mix) improves forget quality at 15% but degrades retain quality, with the authors noting instability.

Limitations

• Focuses only on image-text scenarios; extension to audio, video, or robotics is untested.
• Human-GPT-4o-mini agreement for Forgetfulness is only moderate (ICC 0.57), and the judges disagree in absolute scores.
• Generation-quality gains depend on LLM judges; the paper's mixed-training variant is unstable.

Experimental setup

Datasets: MLLMU-Bench CLEAR DigiFace1M

Benchmarks: MLLMU-Bench CLEAR

Metrics: classification accuracy ROUGE-L contextual refusal score forgetfulness score ICC agreement with humans

Baselines: GA GA_Diff KL_Min NPO MMUnlearner

Builds on: GRPO contrastive activation steering GA_Diff NPO MMUnlearner

Closest papers by meaning

• FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model (NeurIPS 2025 · similarity 0.88)
• Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.88)
• Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
• WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models (NeurIPS 2024 · similarity 0.88)
• RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.88)
• Large Language Model Unlearning (NeurIPS 2024 · similarity 0.87)