Research map / Large language models / Language, safety and interpretability
LLM Unlearning with LLM Beliefs
Kemou Li, Qizhou Wang, Yue Wang, Fengpeng Li, Jun Liu, Bo Han, Jiantao Zhou
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 NPO with BS-S | TOFU unlearning, Llama 3.2 1B · TOFU forget 10% | Agg.: 0.58 -> 0.61 | Improved |
| replace NPO with BS-T | TOFU unlearning, Llama 3.2 1B · TOFU forget 10% | Agg.: 0.58 -> 0.59 | Improved |
| replace NPO with BS-T | TOFU unlearning, Llama 3.1 8B · TOFU forget 10% | Agg.: 0.63 -> 0.63 | No clear effect |
| replace NPO with BS-S | TOFU unlearning, Llama 3.2 1B · TOFU forget 1% | Agg.: 0.53 -> 0.57 | Improved |
| replace NPO with BS-S | TOFU unlearning, Llama 3.2 3B · TOFU forget 1% | Agg.: 0.45 -> 0.50 | Improved |
| replace NPO with BS-S | TOFU unlearning, Llama 3.1 8B · TOFU forget 1% | Agg.: 0.44 -> 0.49 | Improved |
| replace NPO with BS-S | TOFU unlearning, Llama 3.1 8B · TOFU forget 5% | Agg.: 0.53 -> 0.60 | Improved |
| add sequence-level bootstrapping on top of BS-T (BS-S vs BS-T) (ablation) | BS-T · TOFU forget 5%, Llama 3.2 3B | Agg.: 0.55 -> 0.60 | Improved |
| replace WGA with BS-S | TOFU unlearning, Llama 3.2 1B · TOFU forget 10% | Agg.: 0.53 -> 0.61 | Improved |
| replace NPO with BS-S | Zephyr-7B-beta · WMDP Bio | QA accuracy: 0.27 -> 0.26 | Improved |
| replace RMU with BS-S | Zephyr-7B-beta · MMLU (WMDP retain) | accuracy: 0.55 -> 0.54 | Hurt |
| replace NPO with BS-S | Zephyr-7B-beta · MMLU (WMDP retain) | accuracy: 0.44 -> 0.54 | Improved |
Closest papers by meaning
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.92)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.92)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.92)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.91)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.91)
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (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.
