IAPO preprint reports lower token use for matching agent-system accuracy

Researchers introduced IAPO, a method for training an AI agent coordinator, and reported that its deployed plans matched the strongest comparison system's accuracy using fewer tokens. The arXiv preprint compared fifteen published multi-agent and workflow-search systems using the same 7B executor on long-input mathematics, reading and coding tasks. IAPO trains the coordinator to account for how decisions interact, rather than crediting each decision separately. Removing that interaction term reduced coding accuracy by 7 to 17 points. The reported efficiency gains come from these benchmark comparisons.