KABB: Knowledge-Aware Bayesian Bandits for Dynamic Expert Coordination in Multi-Agent Systems

As scaling large language models faces prohibitive costs, multi-agent systems emerge as Multi-Agent Systems (MAS) (Guo et al., 2024b) offer a a promising alternative, though challenged by promising alternative by coordinating multiple specialized static knowledge assumptions and coordination agents to achieve superior performance compared to individual inefficiencies. We introduce Knowledge-Aware systems while maintaining manageable computational Bayesian Bandits (KABB), a novel framework costs and budgets. Recent advances in MAS have led to that enhances multi-agent system coordination the development of several frameworks. For example, the through semantic understanding and dynamic Mixture of Agents (MoA) (Wang et al., 2024) employs multiple adaptation. The framework features three key LLMs as proposers to iteratively refine responses, with innovations: a three-dimensional knowledge distance a central aggregator delivering the final output. Although model for deep semantic understanding, a MoA has demonstrated robustness and scalability in deployment, dual-adaptation mechanism for continuous expert its computational cost scales linearly with the number optimization, and a knowledge-aware Thompson of agents, and significant redundancy and noise become a Sampling strategy for efficient expert selection.