Research map / Continual learning / Continual fine-tuning of LLMs
Attention Smoothing Is All You Need For Unlearning
Saleh Zare Zade, Xiangyu Zhou, Sijia Liu, Dongxiao Zhu
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 fixed-target or divergence forgetting with self-distillation from attention-smoothed forget-teacher (ASU_KL) | IDK_AP · TOFU forget05 | Forget Efficacy: 60.88 -> 77.84 | Improved |
| ASU_KL vs IDK_AP on utility | IDK_AP · TOFU forget05 | Model Utility: 75.23 -> 74.18 | No clear effect |
| ASU_KL vs NPO_GD | NPO_GD · TOFU forget10 | Model Utility: 56.58 -> 73.27 | Improved |
| ASU_KL on real-world unlearning | best baseline utility · Real-world unlearning | Model Utility: 52.76 -> 55.76 | Improved |
| ASU_KL on MUSE News | NPO_KL · MUSE News | VerbMem on forget: 33.0 -> 8.8 | Improved |
| ASU_KL on MUSE Books | NPO_KL · MUSE Books | VerbMem on forget: 54.4 -> 5.3 | Improved |
| smooth attention only in shallow layers 6-8 instead of all layers (ablation) | ASU teacher · TOFU forget05 | MU / FE: FE 77.84 -> 77.49 (KL) | No clear effect |
| combine ASU with IDK_AP refusal outputs (ablation) | IDK_AP · TOFU forget10 | Forget Efficacy: 61.27 -> 86.77 (KL) | Improved |
| vary attention temperature tau from 2.0 to 2.8 (ablation) | ASU · TOFU forget05 | MU and FE | No clear effect |
| increase attention temperature tau in the forget teacher (ablation) | base model · TOFU factual vs function tokens | negative log-likelihood | Mixed |
| apply ASU on continual unlearning sequences | NPO and IDK · TOFU continual forget05/forget10 | average of MU and FE | Improved |
Closest papers by meaning
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.93)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.91)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.91)
- What to Forget in Unlearning? Forget Set Curation for Language Models (NeurIPS 2026 · similarity 0.91)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · 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.
