Xing-Yuan Chen

TDG-LoRA: Token-Level Dynamic Gating for Mitigating Catastrophic Forgetting

Parameter-efficient fine-tuning (PEFT), particularly Low-Rank Adaptation (LoRA), is widely used to adapt large language models (LLMs) to specialized downstream domains. However, although the pretrained backbone remains frozen, a domain-adapted LoRA branch may interfere with the model’s original representations and pred...