Random Two-Peer Routing Beats Learned Topologies on Cost-Accuracy in Sparse Multi-Agent Debate
arXiv·low signal
arXiv 2609.27150 (22 Sep) finds that having each debating agent talk to two distinct, freshly sampled peers per round consistently improves the accuracy-cost trade-off of sparse multi-agent debate. Adding lightweight stopping cuts inference cost further with competitive accuracy. The authors argue that learned or adaptive topology methods should be benchmarked against this simple baseline before their complexity is accepted.