Model Hemorrhage and the Robustness Limits of Large Language Models

Ziyang Ma, Zuchao Li, Lefei Zhang +4

Large language models (LLMs) demonstrate strong performance across natural language processing tasks, yet undergo significant performance degradation when modified for deployment t…