PRISM Protocol Optimizes Permutation Search Strategies with Landscape Diagnostics

This paper introduces PRISM, a novel predictive protocol designed to optimize permutation-based problems by first analyzing the "fitness landscape" of potential orderings. Permutation optimization is relevant whenever the sequence of fixed components impacts overall system performance. Instead of blindly applying search algorithms, PRISM uses low-cost landscape diagnostics, such as one-step move autocorrelation and fitness-distance correlation, to make informed decisions. The protocol can predict which mutation operators will be most effective, identify scenarios where a structured search will outperform random sampling, and even detect when the benefits of any search are minimal. The researchers applied PRISM across diverse domains, including synthetic landscapes, neural architecture benchmarks, scientific machine learning pipelines, and large language model instruction ordering. Experiments revealed that simply reordering instructions in large language models can lead to substantial performance differences, even after prompt wording is optimized, suggesting that content and ordering optimization are complementary. PRISM is not presented as a universally superior optimizer but as a strategic framework to determine the utility of permutation search, the best representation and operator to use, and when simpler alternatives are sufficient.