HyperRepair Recovers General LLM Capabilities After Fine-Tuning

HyperRepair uses a hypernetwork to mitigate catastrophic forgetting in fine-tuned models without increasing inference latency. By learning from activation fingerprints, the framework generates a Repairing Side Net (RSN) that selectively restores general capabilities while maintaining domain specialization.