New Network Improves Continual LLM Unlearning and Capacity Recovery

Machine unlearning aims to completely remove the influence of specific data from a model, a critical capability for privacy and compliance. However, real-world unlearning requests often arrive continuously, posing two significant challenges. First, previous unlearning interventions might cause forgotten knowledge to reappear by redistributing computations across the model's remaining pathways. Second, repeated unlearning can progressively diminish the model's overall capacity, impacting its ability to retain general utility. To tackle these issues, researchers propose the Trajectory-guided Forget-Recover Network (TFR-Net). This network monitors channel-level risk across unlearning requests, distinguishing between persistent target-related channels and temporary "hotspots." It selectively suppresses only the persistent channels, preventing knowledge re-emergence. Furthermore, TFR-Net actively recovers model capacity by reactivating dormant channels that contribute strongly to the model's retained utility and exhibit low historical or current forget risk. This recovery process is only accepted if the degradation of retained utility stays within a predefined tolerance. Experiments across four datasets demonstrate that TFR-Net consistently achieves a better balance between effective unlearning and preserving model utility compared to existing baselines.