Research map / Continual learning / Continual fine-tuning of LLMs
CLUE: Conflict-guided Localization for LLM Unlearning Framework
Hang Chen, Jiaying Zhu, Xinyu Yang, Wenya Wang
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 conflict-mask stage with forget+retain loss (vs replacing it with full mask) (ablation) | CLUE · WMDP Cyber, SST-2 retain | retain utility: down 0.264 | Improved |
| add forget-neuron fine-tuning stage (ablation) | CLUE · WMDP Cyber, SST-2 retain | forget efficacy: down 0.045 | Improved |
| fine-tune conflict neurons before forget neurons (reverse order) (ablation) | CLUE · WMDP Cyber, SST-2 retain | retain utility: drop 0.192 | Hurt |
| use GA instead of PO as forget loss in both stages (ablation) | CLUE · WMDP Cyber, SST-2 retain | retain utility: drop 0.529 | Hurt |
| use NPO instead of PO as forget loss in both stages (ablation) | CLUE · WMDP Cyber, SST-2 retain | retain utility: drop 0.093 | Hurt |
| use CLUE localization instead of WAGLE | WAGLE · WMDP Cyber, SST-2 retain | retain utility: 0.771 -> 0.91 | Improved |
| localize neurons (localization methods) instead of full fine-tuning with PO | PO · WMDP Cyber, SST-2 retain | retain utility: 0.632 -> 0.731 (DEPN) | Improved |
| use CLUE on WMDP Cyber with Winogrande retain set (forget efficacy vs WAGLE) | WAGLE · WMDP Cyber, Winogrande retain | forget efficacy: 0.702 -> 0.697 | No clear effect |
| use MMLU as retain set instead of specific tasks (ablation) | CLUE · WMDP Cyber | forget efficacy: lower than specific-task retain sets | Hurt |
| increase number of specific retain tasks (ablation) | CLUE · WMDP Cyber | forget efficacy: decreases as tasks increase | Hurt |
| use denser circuit (lower sparsity) (ablation) | CLUE · WMDP Cyber/Bio, PKU-SafeRLHF, SST-2 retain | forget efficacy: increases as sparsity decreases from 0.95 to 0.2 | Improved |
Closest papers by meaning
- Adaptive Localization of Knowledge Negation for Continual LLM Unlearning (ICML 2025 · similarity 0.87)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.87)
- Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs (ICLR 2025 · similarity 0.87)
- LoFiT: Localized Fine-tuning on LLM Representations (NeurIPS 2024 · similarity 0.86)
- RA$^2$: Retain-Anchored Attraction Preserves Forget-Adjacent Utility in LLM Unlearning (NeurIPS 2026 · similarity 0.86)
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.85)
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.
