Large language models (LLMs) are increasingly employed for future prediction, prompting the idea of treating multiple models as a wisdom‑of‑the‑crowd mechanism. Simply enlarging the crowd, however, does not guarantee genuine diversity because models may exhibit redundant behaviors. To address this, we introduce a behavior‑aware framework for constructing diverse LLM crowds. The framework first gathers reasoning traces of each model on a set of independent development tasks, treating these traces as behavioral signatures; it then clusters models using K-means++ based on similarity; finally, it selects a representative model (the medoid) from each cluster to form the prediction crowd. We evaluated 25 LLMs on seven development benchmarks to model behavioral diversity and on two future‑prediction benchmarks to assess crowd performance. Results show that crowd composition can outweigh crowd size: a three‑model medoid crowd derived from behavioral clustering outperforms conventional voting over all 25 models on both prediction benchmarks, while reducing model calls by $88\%$ and inference cost by roughly $80\%$. Further analysis indicates that representative behavioral diversity is more important than merely maximizing diversity.
Blogger's Review: This study highlights that, in the era of massive models, carefully selecting a few models with complementary reasoning paths can be far more cost‑effective than indiscriminately scaling up model numbers, offering a practical guideline for system designers.