Research map / Large language models / Language, safety and interpretability
Less is More: Geometric Unlearning for LLMs with Minimal Data Disclosure
Chenchen Tan, Xinghao Li, Shujie Cui, Youyang Qu, Cunjian Chen, Longxiang Gao
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 |
|---|---|---|---|
| edit the last two layers instead of the first layer (ablation) | GU · ToFU Forget-10 (LLaMA3.2-1B) | Extraction strength: 0.498 -> 0.172 | Improved |
| edit the last two layers instead of the middle layer (ablation) | GU · ToFU Forget-10 (LLaMA3.2-1B) | Extraction strength: 0.487 -> 0.172 | Improved |
| edit last two layers instead of the last one (ablation) | GU · ToFU Forget-10 (LLaMA3.2-1B) | Epochs to convergence: 22 -> 27 | Hurt |
| edit last five layers instead of last two (ablation) | GU · ToFU Forget-10 (LLaMA3.2-1B) | Extraction strength / utility: E.S. 0.172 -> 0.165; M.U. 0.577 -> 0.568; epochs 27 -> 35 | Mixed |
| add fold-back and geometric objective under matched synthetic data (GU vs GradDiff) | GU · ToFU Forget-10 matched synthetic data (LLaMA3.2-1B) | Extraction strength: 0.513 -> 0.172 | Improved |
| train NPO on GU synthetic data instead of GU objective | NPO · ToFU Forget-10 matched synthetic data (LLaMA3.2-1B) | Extraction strength: 0.099 -> 0.627 | Hurt |
| train GA on GU synthetic data instead of original data | GA · ToFU Forget-10 matched synthetic data (LLaMA3.2-1B) | Model utility: 0.000 -> 0.231 (GU 0.577) | Hurt |
| increase PCA rank from 4 to 16 (ablation) | GU · safe-behavior subspace (LLaMA2-7B) | explained variance: 63% and 58% -> 98% on both layers | Improved |
| use larger PCA subspace rank (ablation) | GU · ToFU Forget-10 (LLaMA2-7B) | unlearning vs utility | No clear effect |
| increase synthetic samples from 10 to 30 (ablation) | GU · ToFU Forget-10 | Extraction strength | Improved |
| increase synthetic samples from 30 to 40 (ablation) | GU · ToFU Forget-10 | Model utility | Hurt |
| use GU vs GA/RMU on UnlearnPII | GU · UnlearnPII | Utility vs suppression | Mixed |
Closest papers by meaning
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.90)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.90)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.90)
- Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.90)
- Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs (ICLR 2025 · similarity 0.90)
- To Each (Textual Sequence) Its Own: Improving Memorized-Data Unlearning in Large Language Models (ICML 2024 · 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.
