Rectifying Demonstration Shortcut in In-Context Learning

1 code implementation • 14 Mar 2024 • Joonwon Jang, Sanghwan Jang, Wonbin Kweon, Minjin Jeon, Hwanjo Yu

However, LLMs often rely on their pre-trained semantic priors of demonstrations rather than on the input-label relationships to proceed with ICL prediction.

In-Context Learning