Research map / Continual learning / Continual fine-tuning of LLMs
Redirection for Erasing Memory (REM): Towards a universal unlearning method for corrupted data
Stefan Schoepf, Michael Curtis Mozer, Nicole Elyse Mitchell, Alexandra Brintrup, Georgios Kaissis, Peter Kairouz, Eleni Triantafillou
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 the redirection step 3.1 (train on full data with shared mask for forget set) (ablation) | REM · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 69.82 ± 1.85 -> 73.40 ± 1.43 | Improved |
| add the barrier step 3.2 (NPO on theta_o1 during repair) (ablation) | REM without ETD · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 71.38 ± 1.47 -> 73.40 ± 1.43 | Improved |
| add the barrier step 3.2 (ablation) | REM with ETD · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 73.69 ± 0.69 -> 73.19 ± 0.72 | No clear effect |
| train the original model with ETD (ablation) | REM · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 73.40 ± 1.43 -> 73.19 ± 0.72 | No clear effect |
| use REM instead of NPO alone | NPO · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 56.40 ± 2.50 -> 73.40 ± 1.43 | Improved |
| add ETD pretraining | NPO · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 56.40 ± 2.50 -> 67.10 ± 1.17 | Improved |
| add ETD pretraining | SCRUB · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 42.69 ± 2.33 -> 59.85 ± 2.50 | Improved |
| use 100% capacity model instead of 50% capacity | REM · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 73.40 ± 1.43 -> 59.65 ± 2.99 | Hurt |
| use REM instead of Ascent on ViT SVHN | Ascent · SVHN ViT (50% capacity) | Utility x Healed: 67.14 ± 0.89 -> 73.27 ± 0.32 | Improved |
| use retraining from scratch on retain set | Retrained · CIFAR10 ResNet-9 (50% capacity) | Utility x Healed: 48.52 ± 2.47 vs No UL+ETD 50.11 ± 2.75 | Hurt |
| use REM instead of SCRUB on Imagenette | SCRUB · Imagenette (100 corruptions) | Utility x Healed: 54.52 ± 2.89 -> 60.98 ± 2.02 | Improved |
| lower capacity model or ETD for high regularity corruptions | ResNet-9 · poison trigger (high regularity) | Healed | No clear effect |
Closest papers by meaning
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.90)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (NeurIPS 2026 · similarity 0.89)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.88)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.88)
- Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · similarity 0.88)
- LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty (CVPR 2025 · 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.
