Optimizing Corrector Placement for Multi-Agent Swarm Consensus.

A new study explores the challenge of steering a swarm of unreliable agents towards a correct consensus by strategically deploying a limited number of powerful, expensive "oracle" correctors. The core question is how much to spend and where to place these oracles to achieve optimal correctness. The researchers model the swarm as a graph where each oracle influences a node towards the truth with a strength tied to its cost. The findings indicate that the coherence of the swarm remains submodular, meaning each additional oracle provides diminishing returns, even when oracles vary in strength. This property allows a cost-benefit greedy algorithm to achieve a near-optimal placement strategy within any given budget. Furthermore, the study reveals that the decision of whether to invest in a few strong or many medium-strength oracles depends on the curvature of the cost-quality relationship, which is genuinely task-dependent. For instance, this law was found to be concave for math verification but convex for emergent code tracing, suggesting that the optimal strategy varies significantly with the specific task.