This paper presents a theoretical framework and empirical investigation to characterize when cognitive diversity maximizes ensemble gains across machine learning, collective intelligence, and multi-agent systems. It develops a unified theoretical framework that decomposes ensemble performance into accuracy, bias, variance, and diversity components, revealing that diversity operates as a hidden dimension in the bias-variance trade-off. The research proposes testable hypotheses about optimal diversity levels across different task characteristics and validates these through synthetic benchmarks, real-world datasets, and large language model ensembles.
Key findings
Diversity operates as a hidden dimension in the bias-variance trade-off.
Optimal diversity levels vary across different task characteristics.
The research provides actionable guidelines for ensemble design.
Limitations & open questions
Limited characterization of optimal diversity ranges across task types.
Relative importance of different diversity types for tasks remains unclear.