Consilium: How Teams of LLMs Outperform Solo AI

In a novel approach to improving AI performance, a new framework called Consilium lets multiple large language models (LLMs) collaborate to answer questions. Instead of relying on a single model, Consilium orchestrates a panel of LLMs that deliberate and converge on a consensus, mimicking human committee reasoning. Early tests show that this multi-model strategy can boost accuracy and reduce errors, especially on complex reasoning tasks. The system dynamically assigns roles to different LLMs—such as fact-checker, analyst, or summarizer—and then combines their outputs through a voting mechanism. Researchers found that groups of 3 to 5 models significantly outperformed individual LLMs. Consilium represents a shift from bigger models to better orchestration, potentially making AI more reliable without requiring massive single models.