Research map / Continual learning / Continual fine-tuning of LLMs
Towards Natural Machine Unlearning
Zhengbao He, Tao Li, Xinwen Cheng, Zhehao Huang, Xiaolin Huang
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 |
|---|---|---|---|
| use NatMU (inject remaining-data information into forgetting samples) instead of best baseline | relabeling-based unlearning · Full-class Cattle, ideal, CIFAR-100/20 ResNet-18 | Avg.Gap: 0.42 -> 0.11 | Improved |
| use NatMU instead of best baseline | relabeling-based unlearning · Full-class Sea, practical hyperparameters | Avg.Gap: 1.02 -> 0.35 | Improved |
| use NatMU instead of SCRUB | SCRUB · Sub-class Rocket, ideal hyperparameters | Avg.Gap: 1.29 -> 1.74 | Hurt |
| use NatMU instead of best baseline | unlearning baselines · Sub-class Cattle, ideal hyperparameters | Avg.Gap: 3.28 -> 2.22 | Improved |
| use NatMU instead of best baseline | unlearning baselines · Sub-class Sea, practical hyperparameters | Avg.Gap: 4.40 -> 3.63 | Improved |
| use NatMU instead of NegGrad+ | NegGrad+ · CIFAR-10 VGG16-BN, 1% sample-wise | Avg.Gap: 2.18 -> 1.55 | Improved |
| use NatMU instead of SCRUB | SCRUB · CIFAR-100 ResNet-18, 10% sample-wise | Avg.Gap: 2.28 -> 1.77 | Improved |
| use NatMU instead of AMUN | AMUN · TinyImageNet ResNet-34, 1% sample-wise | Avg.Gap: 4.18 -> 2.18 | Improved |
| switch from ideal to practical (transferred) hyperparameters | NatMU · Class-wise unlearning | Avg.Gap | No clear effect |
| switch from ideal to practical hyperparameters | relabeling baselines such as delta-targeted · Sub-class unlearning | Avg.Gap: 6.47 -> 17.56 | Hurt |
| inject correct remaining-data information into forgetting samples (ablation) | relabeling-based unlearning · Forgetting accuracy curves, CIFAR-100 ResNet-18 | accuracy on (x^f*m, y^r) pairs | Improved |
Closest papers by meaning
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.91)
- What makes unlearning hard and what to do about it (NeurIPS 2024 · similarity 0.91)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.90)
- CoUn: Empowering Machine Unlearning via Contrastive Learning (NeurIPS 2025 · similarity 0.89)
- LoRA Unlearns More and Retains More (Student Abstract) (AAAI 2025 · similarity 0.89)
- LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty (CVPR 2025 · similarity 0.89)
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.
