WILC Framework Boosts LLM Collective Intelligence and Cost-Efficiency

Large language models, despite their advancements, still have individual limitations. This presents an opportunity to combine multiple LLMs strategically to achieve a collective intelligence that surpasses any single model. Current methods for combining models often fix their interaction in advance, overlooking the dynamic nature of problem-solving. This research proposes a new framework called WILC (Wisdom Integration of LLM Crowds), which redefines collective LLM intelligence as a sequential, relay-style process. In WILC, each successive model is chosen specifically to address a bottleneck identified in the previous model's output. The framework is built on two core principles: iterative reflection-and-refinement for diagnosing and improving prior outputs, and complementarity-driven model selection. The selection mechanism uses a dual-gate approach: Prospective Complementarity Fit (PCF) identifies the best-suited model for the current bottleneck, while Posterior Complementarity Gain (PCG) evaluates if the chosen transition improves the solution. Experiments across various benchmarks show WILC outperforms single-model self-refinement, ensemble methods, and query-routing. Notably, WILC achieved performance comparable to GPT-5.2 at approximately seven times lower estimated per-query cost, while also supporting data sovereignty through self-hosted deployment.