Multi-LLM Collaboration

Multi-LLM collaboration studies how multiple language models can work together, and when their combined capabilities are more useful than a single model’s. Common approaches assign agents different roles, route difficult questions to stronger or specialized models, and combine answers through voting, verification, debate, or iterative refinement; some systems also exchange hidden states instead of text. Researchers evaluate these methods on reasoning, coding, question answering, long-horizon tasks, and specialized settings such as medicine, while developing ways to select complementary teammates and measure their contribution. A central tension is that collaboration can improve accuracy or lower cost, but communication and coordination add overhead and can amplify errors, agreement, or adversarial influence; reliable gains depend on the task, team, and quality of information exchange.