Research map / Large language models / Language, safety and interpretability
Not All Tokens Are Meant to Be Forgotten
Xiangyu Zhou, Yao Qiang, Saleh Zare Zade, Douglas Zytko, Prashant Khanduri, Dongxiao Zhu
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 unwanted-information identifier (GPT) to baselines GA, NPO, KTO, SimNPO (ablation) | GA/NPO/KTO/SimNPO · TOFU (1%, 5%, 10% forget) | model utility | Improved |
| add unwanted-information identifier to NPO (NPO-GPT) alone (ablation) | NPO · TOFU forget05 | model utility: 74% utility decline remains | Mixed |
| add Preservation Loss on general words to NPO-GPT (ablation) | NPO-GPT · TOFU forget05 | model utility | Improved |
| replace NPO loss with Logit Preference Loss (TPO) (ablation) | NPO+PL-GPT · TOFU forget05 | model utility and KL divergence | Improved |
| use NPO with the entire forget sample as target (larger forget set) | NPO · TOFU | model utility: 0.76 -> 0.11 | Hurt |
| increase forget set size to 10% | GA and KTO based methods · TOFU forget 10% | forget quality: near-zero forget quality | Hurt |
| use TPO-GPT at 10% forget set | TOFU LLaMA-2 7B and LLaMA-3.2 3B · TOFU forget 10% | forget quality and utility: 70% utility with highest forget quality | Improved |
| add GDR retain loss to unlearning methods (ablation) | GA/NPO/SimNPO/TPO · MUSE News and Books (LLaMA-2 7B) | KnowMem on Dr: TPO 0.00 -> 43.67 (News), 0.00 -> 42.07 (Books) | Improved |
| use TPOGDR instead of SimNPOGDR | MUSE News · MUSE News (LLaMA-2 7B) | PrivLeak: -97.17 -> -6.12 | Improved |
| use Task Vector baseline | Task Vector · MUSE | forgetting: VerbMem 66.74 vs original 56.26 (News) | Hurt |
| use ChatGPT-4 generative identifier instead of DistilBERT discriminative identifier (ablation) | TIF · TOFU | forget quality vs utility balance | Improved |
Closest papers by meaning
- What to Forget in Unlearning? Forget Set Curation for Language Models (NeurIPS 2026 · similarity 0.91)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.91)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.90)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.89)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.89)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 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.
