Why Prompt Optimization Works, and Why It Sometimes Doesn't: A Causal-Inspired Edit-Level Analysis

In three linesCausal analysis of prompt optimization methods (DSpy, TextGrad) explaining generalization failures. Complexity-increasing edits harm mathematical and multi-hop reasoning, while step-by-step edits improve logical reasoning. Failures stem from systematic interactions between edit families and task characteristics, not random artifacts.

Summary generated by Claude — human-verified