PassDrill Free practice questions with explanations that actually teach.

AI-ENGINEERING · prompting · Q007 · medium

Zhao et al. (2021), "Calibrate Before Use: Improving Few-Shot Performance of Language Models," identify "recency bias" as one cause of instability in few-shot prompting. What does recency bias describe?

  1. The model's tendency to disproportionately predict whichever label appeared in the example placed nearest the end of the few-shot prompt, regardless of the true input
  2. The tendency of a model's accuracy to decline over time as new versions of the model are released
  3. The model favoring the most recently published research papers when asked to cite sources
  4. The tendency to give more weight to the very first example in a few-shot prompt while ignoring later examples