Tag
The paper proposes a behavior-aware framework for constructing diverse LLM crowds to improve future prediction, demonstrating that smaller, behaviorally diverse groups can outperform larger ones while reducing inference costs.