Research map / Large language models / Language, safety and interpretability
Distinguishable Deletion: Unifying Knowledge Erasure and Refusal for Large Language Model Unlearning
Puning Yang, Junchi Yu, Qizhou Wang, Philip Torr, Bo Han, Xiuying Chen
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 |
|---|---|---|---|
| replace token-suppression KD methods with energy-bounded unlearning (EUA) | best KD baseline on LLaMA-3.2-3B · TOFU Forget 10% | Agg.: 6.967 -> 8.665 | Improved |
| add energy-based refusal and energy-boundary training over DR baselines | Guardnail on LLaMA-3.1-8B · TOFU Forget 10%, after jailbreak attack | Detection accuracy: 95.6% -> 100.0% | Improved |
| use EUA instead of unmodified model with DR prompting | DR methods on LLaMA-3.1-8B · TOFU Forget 10% | Retention Quality: 7.456 -> 7.433 | Hurt |
| use self-preferenced energy margins instead of manual margins (ablation) | EUA on LLaMA-3.2-3B · TOFU Forget 10% | Agg.: 7.944 -> 8.665 | Improved |
| use self-preferenced energy margins instead of manual margins (ablation) | EUA on Qwen2.5-3B · TOFU Forget 10% | Erasing Quality: 7.652 -> 9.528 | Improved |
| vary top-k used for energy and threshold at inference (ablation) | EUA · TOFU | decision reliability | No clear effect |
| add EUA over SatImp | SatImp on MUSE-Books · MUSE-Books (ICLM-7B) | UtilPres: 45.77 -> 50.45 | Improved |
| use EUA instead of SatImp for Bio unlearning | SatImp on Zephyr-7B-beta · WMDP Bio | accuracy (lower is better): 0.2545 -> 0.2567 | No clear effect |
| EUA under cross-lingual, relearning and jailbreak attacks | EUA on LLaMA-3.2-3B · TOFU | EQ and LQ | No clear effect |
Closest papers by meaning
- SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.90)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.90)
- Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs (ICLR 2025 · similarity 0.90)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.90)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.90)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · 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.
