Research map / Large language models / Language, safety and interpretability
Randomized Antipodal Search Done Right for Data Pareto Improvement of LLM Unlearning
Ziwen Liu, Huawei Lin, Yide Ran, Denghui Zhang, Jianwen Xie, Chuan Li, Weijie Zhao, Zhaozhuo Xu
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 randomized paired retain-set selection (RASLIK vs RASLIK-F) (ablation) | RASLIK · Howdy-Alpaca, OLMo-2-1124-7B, GA_GDR | D_mah: 10.660 -> 9.813 | Improved |
| add randomized paired retain-set selection (RASLIK vs RASLIK-F) (ablation) | RASLIK · Howdy-Alpaca, OLMo-2-1124-7B, GA_GDR | Non-SF: 0.466 -> 0.911 | Improved |
| add randomized paired retain-set selection (RASLIK vs RASLIK-F) (ablation) | RASLIK · Virtual-Alpaca, Pythia-2.8B, GA_KLR | D_mah: 29.150 -> 27.670 | Improved |
| add randomized paired retain-set selection (RASLIK vs RASLIK-F) (ablation) | RASLIK · Virtual-Alpaca, OLMo-2-1124-7B, GA_GDR | D_mah: 87.333 -> 87.166 | Improved |
| use RASLIK instead of Random Selection | unlearning data retrieval · Howdy-Alpaca, OLMo-2-1124-7B, GA_GDR | Non-SF: 0.040 -> 0.911 | Improved |
| use RASLIK instead of Oracle Sampling | unlearning data retrieval · Howdy-Alpaca, OLMo-2-1124-7B, GA_GDR | D_mah: 11.083 -> 9.813 | Improved |
| use RASLIK instead of Oracle Sampling | unlearning data retrieval · Howdy-Alpaca, Pythia-2.8B, GA_GDR | D_mah: 38.081 -> 38.622 | Hurt |
| use RASLIK instead of Oracle Sampling | unlearning data retrieval · Virtual-Alpaca, Pythia-2.8B, GA_GDR | D_mah: 56.113 -> 55.458 | Improved |
| mix 62% Oracle with 38% random forget samples (CR-62) instead of pure Oracle (ablation) | Oracle Sampling · Howdy-Alpaca, Pythia-2.8B, GA_GDR | D_mah: 56.331 -> 55.139 | Improved |
| mix 50% Oracle with random forget samples (CR-50) instead of pure Oracle (ablation) | Oracle Sampling · Howdy-Alpaca, Pythia-2.8B, GA_KLR | D_mah: 107.746 -> 107.207 | Improved |
| mix 25% Oracle with random forget samples (CR-25) instead of pure Oracle (ablation) | Oracle Sampling · Howdy-Alpaca, Pythia-2.8B, GA_GDR | D_mah: 56.331 -> 56.936 | Hurt |
| use paired RASLIK noise instead of CR-62 unstructured mixing (ablation) | CR-x variants · Howdy-Alpaca, Pythia-2.8B, GA_GDR | D_mah: 55.139 -> 54.989 | Improved |
Closest papers by meaning
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.90)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.89)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.89)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.89)
- SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.89)
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (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.
