Research map / Large language models / Language, safety and interpretability
Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding
Hadi Reisizadeh, Jiajun Ruan, Yiwei Chen, Soumyadeep Pal, Sijia Liu, Mingyi Hong
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 |
|---|---|---|---|
| evaluate with probabilistic decoding (sampling, k=128) instead of greedy decoding | unlearned LLMs (NPO, GradDiff, BLUR-NPO) · TOFU and MUSE-News | Leak@k | Hurt |
| raise top-p from 0.2 to 1.0 at T=1.0, k=128 (ablation) | NPO · TOFU | Leak@k (ES) | Hurt |
| raise temperature while keeping top-p low (0.2) (ablation) | unlearned models · TOFU | Leak@k (ES) | No clear effect |
| raise top-p while keeping temperature low (ablation) | unlearned models · TOFU | Leak@k (ES) | Hurt |
| increase number of generations k from 1 to 128 at (T,p)=(0.8,1.0) (ablation) | all unlearned methods · MUSE-News | Leak@k (RS) | Hurt |
| apply NPO unlearning, evaluated with Leak@k accuracy | Zephyr-7B-beta · WMDP-bio | Leak@k Acc: 26% | No clear effect |
| apply RMU unlearning under high-randomness sampling (T,p)=(1.0,1.0) | Zephyr-7B-beta · WMDP-bio | Leak@k (LJ) | Hurt |
| add RULE (self-generated leaking responses added to forget set) | GradDiff · TOFU | Leak@k (ES): RULE-GradDiff achieves almost ES=0 across all k | Improved |
Closest papers by meaning
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.92)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.91)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.91)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.91)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 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.
