Individually harmless prompt constraints combine into a 12-point accuracy drop, and only on some model families

arXiv 2609.03156·medium signalXBlueskyLinkedInCopy link

arXiv 2609.03156 runs a full-factorial 3x3x3 design over output format, persona assignment and urgency framing across all 164 HumanEval+ problems and five OpenAI models, yielding 22,140 greedy-decoding evaluations, and decomposes each compound condition into an additive prediction plus a residual interaction term. The GPT-4o family shows consistent super-additive degradation of 3-12 percentage points beyond the additive prediction, the largest being -12.2 pp on GPT-4o-mini for JSON plus expert persona plus moderate urgency; JSON interacts worse than XML, the GPT-4.1 family is largely resistant, and o3-mini improves under structured output constraints. Vulnerability tracked architecture, not size, so a production prompt that stacks constraints needs compound testing rather than single-factor ablations.