Research map / Large language models / Language, safety and interpretability
Dual-Space Smoothness for Robust and Balanced LLM Unlearning
Han Yan, Zheyuan Liu, Meng Jiang
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 parameter-space smoothness (ablation) | PRISM · MUSE-Books relearning attack 100 steps | VerbMem on forget set: 16.664 -> 6.804 | Improved |
| add representation-space smoothness (robust probe) (ablation) | PRISM · MUSE-Books relearning attack 100 steps | VerbMem on forget set: 11.112 -> 6.804 | Improved |
| add gradient-conflict decoupling (ablation) | PRISM · MUSE-Books relearning attack 50 steps | utility on retain set: 1.333 -> 46.588 | Improved |
| add representation-space smoothing (ablation) | PRISM · MUSE-Books relearning attack 50 steps | utility on retain set: 24.412 -> 46.588 | Improved |
| add parameter-space smoothing (ablation) | PRISM · MUSE-Books relearning attack 50 steps | utility on retain set: 13.506 -> 46.588 | Improved |
| use PRISM instead of SAM+NPO | PRISM · MUSE-News | Unlearn Score: 0.000 -> 0.522 | Improved |
| use PRISM instead of RMU-LAT under Relearn-25% at 50 steps | PRISM · MUSE-Books Relearn-25% | VerbMem on forget set: 93.87 -> 0.082 | Improved |
| add PRISM to the unlearned model | PRISM · MUSE-Books | representation margin: median margin +24.9% | Improved |
| use PRISM instead of GDR-regularized baselines | PRISM · WMDP XStest | refusal rate: 0.880 -> 1.000 | Hurt |
| use PRISM instead of SAM+NPO | PRISM · MUSE-Books | time per step (s): 11.055 -> 11.223 | Hurt |
| use PRISM instead of NPO_GDR | PRISM · WMDP multi-turn jailbreak | ASR: 0.326 -> 0.196 | Improved |
Closest papers by meaning
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.91)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.91)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.90)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.89)
- Textual Unlearning Gives a False Sense of Unlearning (ICML 2025 · similarity 0.89)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.89)
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.
