基于语义路由的可逆终身模型编辑LoRA

The dynamic evolution of the real world necessitates model editing within Large Language Models. Existing methods, while exploring modular isolation or parameter-efficient strategies, still suffer from semantic drift or knowledge forgetting due to continual updates. To address these challenges, a Semantic routing-based LoRA framework, SoLA, is proposed for lifelong model editing. In SoLA, each edit is encapsulated as an independent LoRA module, which is frozen after training and mapped to input via semantic routing, allowing for dynamic activation of LoRA modules through semantic matching. This mechanism prevents semantic drift caused by cluster updates and mitigates catastrophic forgetting from parameter sharing. Importantly, SoLA supports precise model editing, enhancing the model's flexibility and adaptability.