Automatic Prompt Optimization

Automatic prompt optimization studies how to improve natural-language instructions, examples, and sometimes prompt structure without changing model weights. Common approaches use black-box evaluation with evolutionary search, textual-gradient or reflective edits, Bayesian and bandit selection, and targeted example selection; newer systems also optimize multi-step workflows, decoding settings, and task-specific prompt components. Work increasingly treats evaluation budget, inference cost, interpretability, personalization, and stability as objectives alongside accuracy, using held-out tests and multi-objective selection to limit regressions. A central tension is that gains can depend on the task and model: noisy or narrow feedback may cause overfitting, prompt hacking, or regressions, so reliable transfer matters as much as peak benchmark scores.