DualOptim+: Bridging Shared and Decoupled Optimizer States for Better Machine Unlearning in Large Language Models
About
We propose DualOptim+, a novel optimization framework for improving machine unlearning in large language models. It introduces a base state to capture common representations shared by forgetting and retaining objectives and delta states to preserve objective-specific residuals. This architecture allows the optimizer to adaptively bridge shared and decoupled states based on the directional conflict between forgetting and retaining gradients. We further introduce DualOptim+ 8bit, a quantized variant that reduces memory overhead without compromising performance. Extensive experiments across fictitious and real-world unlearning, safety alignment, and multi-task learning tasks demonstrate that DualOptim+ consistently achieves a superior trade-off between different objectives. Codes are available at https://github.com/CityU-MLO/DualOptimPlus.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Machine Unlearning | TOFU Forget05 Phi-1.5B model (5%) | Model Utility (MU)51.52 | 32 | |
| Machine Unlearning | TOFU (forget 1% data) | UFE97.88 | 24 | |
| Machine Unlearning | TOFU (forget 5% data) | UFE96.91 | 24 | |
| Machine Unlearning | TOFU (forget 10% data) | UFE Score96.85 | 24 | |
| Unlearning | TOFU Phi 1.5 1.0 (forget 10%) | UFE66.88 | 12 | |
| Unlearning | TOFU Llama 2 1.0 (forget 1%) | UFE89.47 | 12 | |
| Unlearning | TOFU Llama 2 1.0 (forget 5%) | UFE Score84.49 | 12 | |
| Unlearning | TOFU Llama 2 1.0 (forget 10%) | UFE82.18 | 12 | |
| Unlearning | TOFU Phi 1.5 1.0 (forget 1%) | UFE75.51 | 12 | |
| Model Utility | Alpaca-Llama 3 Utility Benchmarks | ARC-c Score47.27 | 7 |