When More Examples Make Your LLM Worse: Discovering Few-Shot Collapse

I found that adding few-shot examples can actively degrade LLM performance. Here are three patterns I discovered and an open-source tool to detect them.