Research map / Vision-language and multimodal / Multimodal LLMs
Beyond Sample-Level Forgetting: Improving Reliability in Multimodal Unlearning
Jianzhou Wang, Yirui Wu, Lixin Yuan, Wenxiao Zhang, Jun Liu
Read the paperCheck your idea against 137,217 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 a causal-perspective approach to multimodal unlearning that decouples unimodal and multimodal knowledge to improve effectiveness, reliability, and locality.
Problem
Multimodal unlearning, removing specific data from pretrained multimodal models, struggles to achieve effectiveness, reliability, and locality because unimodal and multimodal knowledge are interdependent.
Approach
Introduces a causal perspective and performs multimodal unlearning with decoupled unimodal and multimodal knowledge, moving beyond sample-level forgetting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add negative pairs (N-Pairs) (ablation) | contrastive semantic editing · retrieval privacy 50 samples | Meff: 95.82 -> 99.54 | Improved |
| add negative pairs (N-Pairs) (ablation) | contrastive semantic editing · retrieval privacy 50 samples | Mrel: 96.3 -> 99.33 | Improved |
| add positive pairs (P-Pairs) (ablation) | contrastive semantic editing · retrieval privacy 50 samples | Meff / Mrel: Meff 86.16 -> 99.54; Mrel 83.79 -> 99.33 | Improved |
| add gradient-based optimization (ablation) | contrastive semantic editing · retrieval privacy 50 samples | Mloc: 66.38 -> 86.09 | Improved |
| generate Rj with both U'i and K'i instead of only U'i (ablation) | generation strategy · retrieval privacy 50 samples | Mrel: 76.60 -> 99.33 | Improved |
| generate Rj with both U'i and K'i instead of only K'i (ablation) | generation strategy · retrieval privacy 50 samples | Mloc: 64.22 -> 86.09 | Improved |
| swap in proper U'i (vs [U'j, K'i]) (ablation) | generation strategy · retrieval privacy 50 samples | Mloc: 62.57 -> 86.09 | Improved |
| use structured generated Rj instead of random selecting Xj (ablation) | generation strategy · retrieval privacy 50 samples | Meff / Mrel: Meff 98.72 -> 99.54; Mrel 85.35 -> 99.33 | Improved |
| random selecting Xj as Rj instead of generated (ablation) | generation strategy · retrieval privacy 50 samples | Mloc: 89.7 vs 86.09 | Mixed |
| increase number of generated images NR1 (ablation) | virtual samples · retrieval privacy 50 samples | reliability and locality | Hurt |
| increase number of generated texts NR2 (ablation) | virtual samples · retrieval privacy 50 samples | performance | No clear effect |
| use adaptive token-wise alpha instead of fixed alpha (ablation) | contrastive semantic editing · retrieval privacy 50 samples | Meff / Mrel / Mloc | Improved |
Closest papers by meaning
- SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning (CVPR 2026 · similarity 0.90)
- Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models (AAAI 2026 · similarity 0.90)
- Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models (NeurIPS 2025 · similarity 0.88)
- Towards Unified Multimodal Editing with Enhanced Knowledge Collaboration (NeurIPS 2024 · similarity 0.88)
- Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
