PMPO: Probabilistic Metric Prompt Optimization for Small and Large Language Models

원문 초록

Prompt optimization offers a practical and broadly applicable alternative to fine-tuning for improving large language model (LLM) performance. However, existing methods often rely on costly output generation, self-critiquing abilities, or human-annotated preferences, which limit their scalability, especially for smaller or non-instruction-tuned models. We introduce PMPO (Probabilistic Metric Prompt Optimization), a unified framework that refines prompts using token-level cross-entropy loss as a direct, lightweight evaluation signal. PMPO identifies low-quality prompt segments by masking and measuring their impact on loss, then rewrites and selects improved variants by minimizing loss over positive and negative examples. Unlike prior methods, it requires no output sampling or human evaluation during optimization, relying only on forward passes and log-likelihoods. PMPO supports both supervised and preference-based tasks through a closely aligned loss-based evaluation strategy. Experiments show that PMPO consistently outperforms prior methods across model sizes and tasks: it achieves the highest average accuracy on BBH, performs strongly on GSM8K and AQUA-RAT, and improves AlpacaEval 2.0 win rates by over 19 points. These results highlight PMPO's effectiveness, efficiency, and broad applicability.