EMS: Multi-Agent Voting via Efficient Majority-then-Stopping

The article presents Efficient Majority-then-Stopping (EMS), a novel framework for multi-agent voting that enhances computational efficiency by implementing a reliability-aware agent scheduling approach. EMS utilizes a Task-Conditioned Reliability Ordering (TCRO) to prioritize agents based on historical consensus data, and employs Adaptive Incremental Voting (AIV) to terminate the voting process early when a majority consensus is reached, reducing the average number of invoked agents by 35% and token consumption by 44% across five benchmarks. This method is significant for practitioners as it maintains the accuracy of traditional majority voting while optimizing resource usage in multi-agent systems.