DualOptim+ Framework Achieves Superior Unlearning Trade-offs in LLMs
New optimizer bridges shared and decoupled states for 8-bit efficient unlearning
DualOptim+ introduces a novel optimization architecture for machine unlearning in large language models. It maintains a base state capturing common representations shared by forgetting and retaining objectives, along with delta states that preserve objective-specific residuals. This design allows the optimizer to adaptively bridge shared and decoupled states based on the directional conflict between forgetting and retaining gradients. The framework also includes DualOptim+ 8-bit, a quantized version that significantly reduces memory overhead without compromising performance, making it practical for resource-constrained environments.
Extensive experiments demonstrate DualOptim+'s effectiveness across diverse tasks: synthetic and real-world unlearning, safety alignment, and multi-task learning. The method consistently achieves a better trade-off between forgetting the target data and retaining model utility compared to baselines. Accepted at ICML 2026, the code is publicly available, enabling further research into efficient and safe LLM modification. This work addresses a growing need for practical unlearning mechanisms in deployed LLMs, with implications for compliance and model safety.
- DualOptim+ uses a shared base state and decoupled delta states to balance forgetting and retaining objectives adaptively.
- An 8-bit quantized variant reduces memory overhead with no performance loss, enabling practical deployment.
- Validated across synthetic/real-world unlearning, safety alignment, and multi-task learning; accepted at ICML 2026.
Why It Matters
Enables practical, memory-efficient machine unlearning for LLMs, critical for compliance, safety, and dynamic model updating.