Research map / Large language models / Language, safety and interpretability
Revisiting the Past: Data Unlearning with Model State History
Keivan Rezaei, Mehrdad Saberi, Abhilasha Ravichander, Soheil Feizi
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 checkpoint close to forget set (3691B/3859B/last vs 500B) (ablation) | MSA · TOFU forget10 (OLMo-2-7B) | Acc_retain: 0.64 -> 0.83 | Improved |
| apply MSA with pre-exposure checkpoint instead of Task Vector from target model (ablation) | MSA · TOFU forget10 (OLMo-2-7B) | Acc_recover: 0.26 -> 0.42 | Improved |
| MSA vs RMU on recovering ideal-model behavior | MSA · TOFU forget10 (OLMo-2-7B) | Acc_recover: 0.08 -> 0.42 | Improved |
| MSA vs NPO on RESTOR | MSA_instruct · RESTOR Llama-3.1-8B-Instruct | accuracy: 48.45 -> 63.95 | Improved |
| MSA vs best baseline on RESTOR OLMo-2-7B | MSA_last · RESTOR OLMo-2-7B | accuracy: 40.25 -> 47.80 | Improved |
| use earlier checkpoint (500B instead of last) (ablation) | MSA · RESTOR OLMo-2-7B | accuracy: 47.80 -> 45.67 | Mixed |
| use earlier checkpoint (500B instead of last) (ablation) | MSA · MUSE-Books KnowMem retain | KnowMem: 0.65 -> 0.51 | Hurt |
| MSA vs baselines on membership inference | MSA_last · MUSE-Books | PrivLeak: 14.67 vs -99.93 (NPO), -17.83 (RMU) | Improved |
| use checkpoint before forget-set exposure rather than checkpoint after TOFU when forget set is not last (ablation) | MSA · TOFU forget10 + C4 (Llama-3.2-1B) | Acc_recover: 0.37 -> 0.45 | Improved |
| base vs instruct checkpoint for forget vector (ablation) | MSA · TOFU forget10 (Llama-3.1-8B-Instruct) | Acc_recover: 0.45 -> 0.46 | No clear effect |
Closest papers by meaning
- MUSE: Machine Unlearning Six-Way Evaluation for Language Models (ICLR 2025 · similarity 0.92)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.91)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.91)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.91)
- MLUBench: A Benchmark for Lifelong Unlearning Evaluation in MLLMs (ICML 2026 · similarity 0.90)
- Large Language Model Unlearning (NeurIPS 2024 · 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.
