Difficulty-aware topology selection beats always-hierarchical multi-agent coding by 4.1 points at 40% of the cost

arXiv·medium signalXBlueskyLinkedInCopy link

A September 12 arXiv paper measured five communication topologies across 614 problems from APPS, HumanEval+ and LiveCodeBench, finding hierarchical collaboration worth only 2.4 pass@1 points on the easiest problems but 21.1 points on the hardest, at a flat ~10x token cost throughout. Its DATS selector predicts each topology's success probability and picks the one maximizing predicted success minus cost, using a graph network that treats topologies as ordered nodes (worth 1.7 points over a flat multi-label head). Held to 40% of always-hierarchical spend, DATS reached 77.7% pass@1 against 73.6% for always-hierarchical and 74.3% for the strongest learned competitor, holding across four model variants spanning 14 capability points.