Prompt Optimizers Cause Requirement Regressions in Unscored Constraints

A study of GEPA, MIPROv2, and TextGrad shows that optimizing for single metrics often degrades unscored requirements like citation validity or JSON formatting. While regressions in citation accuracy reach up to 41.6 points on HotpotQA, post-hoc analysis suggests much of this is due to output length constraints causing truncated responses.