Research map / Continual learning / Continual fine-tuning of LLMs
DualOptim+: Bridging Shared and Decoupled Optimizer States for Better Machine Unlearning in Large Language Models
Xuyang Zhong, Qizhang Li, Steven Y. Guo, Chen Liu
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 shared base state with decoupled delta states | DualOptim · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 57.96 -> 59.57 | Improved |
| add shared base state with decoupled delta states | DualOptim · Real-world unlearning ME+GD (Llama-3-8B) | OVR: 67.60 -> 73.13 | Improved |
| add shared base state with decoupled delta states | DualOptim · Real-world unlearning IDK+GD (Llama-3-8B) | OVR: 52.73 -> 52.90 | Mixed |
| add shared base state with decoupled delta states | DualOptim 8bit · Multi-task learning (Llama-2-7B) | average accuracy: 64.36 -> 67.14 | Improved |
| use alternating updates instead of joint update | Joint · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 50.85 -> 56.91 | Improved |
| update base state after parameter update instead of before delta (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 58.93 -> 59.57 | Improved |
| update base state after parameter update instead of after delta (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 58.14 -> 59.57 | Improved |
| update base state with raw gradient instead of g minus delta (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 59.08 -> 59.57 | Improved |
| use fast momentum for both states instead of slow for both (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 57.59 -> 59.57 | Improved |
| use fast momentum for base state instead of slow (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 58.88 -> 59.57 | Improved |
| quantize base and delta states to 8 bit (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 59.57 -> 58.55 | Hurt |
| set retain frequency Fr=5 instead of 1 (ablation) | DualOptim+ · TOFU forget 5% IDK+GD (Phi-1.5) | OVR: 59.13 -> 61.30 | Improved |
Closest papers by meaning
- Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models (NeurIPS 2025 · similarity 0.92)
- DualOptim: Enhancing Efficacy and Stability in Machine Unlearning with Dual Optimizers (NeurIPS 2025 · similarity 0.91)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
- Large Language Model Unlearning via Embedding-Corrupted Prompts (NeurIPS 2024 · similarity 0.89)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.89)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.89)
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.
