I Tested 12 LLMs With Few-Shot Examples. The Results Were Not What I Expected.

In a previous article, I tested 8 models across 4 tasks and reported on "few-shot collapse" — cases...