random-text-control

This paper challenges the distortion hypothesis for few-shot degradation in language models by introducing a random-text control, showing that representation shift is largely due to prompt length, and models with higher content delta benefit more from few-shot prompting.