Research map / Continual learning / Continual fine-tuning of LLMs
Towards Benign Memory Forgetting for Selective Multimodal Large Language Model Unlearning
Zhen Zeng, Leijiang Gu, Zhangling Duan, Feng Li, Zenglin Shi, Cees Snoek, Meng Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retaining-anchor masking (directional conflict and relative magnitude) to MFA (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Retain-set image memory ROUGE-L: 0.486 -> 0.622 | Improved |
| add retaining-anchor masking to MFA (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 10% forget) | Retain-set image understanding ROUGE-L: 0.530 -> 0.634 | Improved |
| add retaining-anchor masking to MFA (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 10% forget) | Forget-set image memory ROUGE-L: 0.198 -> 0.493 | Hurt |
| add retaining-anchor masking to MFA (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Forget-set image understanding ROUGE-L: 0.629 -> 0.655 | Improved |
| add relative magnitude mask to directional conflict mask (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Forget-set image memory ROUGE-L: 0.641 -> 0.460 | Improved |
| add directional conflict mask to relative magnitude mask (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Forget-set image memory ROUGE-L: 0.637 -> 0.460 | Improved |
| add relative magnitude mask to directional conflict mask (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 10% forget) | Forget-set text memory ROUGE-L: 0.744 -> 0.566 | Improved |
| add both masks versus single mask (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Retain-set image memory ROUGE-L: 0.682 -> 0.622 | Hurt |
| increase masking hyperparameter k (ablation) | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) and 10% | forget-set ROUGE-L | Improved |
| SMFA vs IDK Tuning under LLaVA 5% overall memory ROUGE-L | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Overall memory ROUGE-L: 0.121 -> 0.199 | Improved |
| SMFA vs IDK Tuning under full fine-tuning, retain image understanding | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) (full fine-tuning) | Retain-set image understanding ROUGE-L: 0.105 -> 0.640 | Improved |
| apply Distractors prompt attack to SMFA | SMFA · S-MLLMUn Bench (LLaVA-OneVision, 5% forget) | Forget-set image memory ROUGE-L: 0.460 -> 0.385 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Model Tailor — Model Tailor: Mitigating Catastrophic Forgetting in Multi-modal Large Language Models (ICML 2024) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models (NeurIPS 2024 · similarity 0.92)
- Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
- Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models (AAAI 2026 · similarity 0.90)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.90)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.90)
- Selective Forgetting: Advancing Machine Unlearning Techniques and Evaluation in Language Models (AAAI 2025 · similarity 0.90)
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.
