Research map / Architectures and efficiency / RNNs and normalization
LMCleaner: Efficient and Certified Online Unlearning via Influence Propagation Truncation
Jie Xu, Zihan Wu, Wenbo Pan, Jiao Yin, Yong-Feng Ge, Hua Wang, Cong Wang, Xiaohua Jia
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 |
|---|---|---|---|
| increase truncation window K from 50 to 1000 (ablation) | LMCleaner · TOFU Epoch 5, LLaMA-3.2-1B | utility, forget truth, MIA: utility 0.429 -> 0.428; 20x cost | No clear effect |
| increase truncation window K to 1250 (ablation) | LMCleaner · TOFU Epoch 5, LLaMA-3.2-1B | MIA: 0.516 -> 0.599 | Hurt |
| increase mini-batch size B from 8 to 256 (ablation) | LMCleaner · TOFU, LLaMA-3.2-1B | FG Prob / utility: FG Prob 0.282 -> 0.201; utility 0.388 -> 0.349 | Mixed |
| increase batch size B from 8 to 64 (ablation) | LMCleaner · TOFU, LLaMA-3.2-1B | MIA: 0.575 -> 0.509 | Improved |
| treat mini-batch as atomic unit instead of per-sample gradients | LMCleaner · storage | storage reduction: 512x reduction | Improved |
| use online unlearning during training instead of fine-tuning-based unlearning | LMCleaner vs GradDiff · TOFU Epoch 5, LLaMA-3.2-1B | model utility: 0.270 -> 0.429 | Improved |
| LMCleaner vs PDU on utility at 3B | LLaMA-3.2-3B · TOFU Epoch 1 | model utility: 0.439 (PDU) vs 0.431 | No clear effect |
| use LMCleaner for membership inference resistance | LLaMA-3.2-3B · TOFU Epoch 1 | MIA min-k: 0.287 (PDU) -> 0.332 | Improved |
| use LMCleaner instead of NPO over epochs | GPT-2 · TOFU Epoch 4 | utility stability | Improved |
Closest papers by meaning
- A Certified Unlearning Approach without Access to Source Data (ICML 2025 · similarity 0.87)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.87)
- Rewind-to-Delete: Certified Machine Unlearning for Nonconvex Functions (NeurIPS 2025 · similarity 0.87)
- The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning (ICLR 2025 · similarity 0.87)
- Certified Machine Unlearning via Noisy Stochastic Gradient Descent (NeurIPS 2024 · similarity 0.87)
- Efficient Source-free Unlearning via Energy-Guided Data Synthesis and Discrimination-Aware Multitask Optimization (ICML 2025 · similarity 0.87)
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.
