Research map / Large language models / Language, safety and interpretability
Explainable LLM Unlearning through Reasoning
Junfeng Liao, Qizhou Wang, Shanshan Ye, Xin Yu, Ling Chen, Zhen Fang
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 GA-only unlearning with reasoning-based targets (TRU) | Zephyr-7B-beta · WMDP-Bio | Relevance (UQ): 0.00 -> 6.72 vs GradDiff | Improved |
| use TRU instead of RMU | Zephyr-7B-beta · WMDP-Bio | unlearning quality: 2.89 -> 6.72 (Rel) | Improved |
| use TRU instead of RMU | Zephyr-7B-beta · WMDP-Bio | retention specificity: 6.42 -> 2.56 | Hurt |
| add reasoning traces to the unlearning target (ablation) | TRU · WMDP-Bio | retention specificity: 0.15 -> 2.56 | Improved |
| add reasoning traces to the unlearning target (ablation) | TRU · TOFU-Forget05 | retention specificity: 0.32 -> 3.68 | Improved |
| add target loss L_target (ablation) | TRU · WMDP-Bio and TOFU-Forget05 | UQ and RQ | Improved |
| add GA-based loss (ablation) | TRU · WMDP-Bio and TOFU-Forget05 | UQ and RQ | Improved |
| add well-specified target criteria (ablation) | TRU · WMDP-Bio and TOFU-Forget05 | UQ and RQ | Improved |
| enlarge unlearning scope from author profile to personal information (ablation) | TRU · TOFU-Forget01 | specificity: 4.23 -> 2.31 | Hurt |
| apply cross-lingual (Spanish, Russian) attack | TRU · WMDP-Bio | UQ: -0.24 (Spanish), -0.47 (Russian) | No clear effect |
| apply jailbreak prompts | TRU · WMDP-Bio | UQ: -0.33 and -0.65 | No clear effect |
| apply relearning attacks | TRU · WMDP-Bio | UQ: 7.01 -> 6.56 and 7.01 -> 6.62 | No clear effect |
Closest papers by meaning
- Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs (ICLR 2025 · similarity 0.93)
- Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.92)
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (NeurIPS 2025 · similarity 0.91)
- Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning (ICML 2025 · similarity 0.91)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.91)
- Adaptive Localization of Knowledge Negation for Continual LLM Unlearning (ICML 2025 · similarity 0.91)
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.
