A position-paper debate rendered as structure: when does explicit decomposition, routing, and graph orchestration beat a single giant model? This page treats the episode as a systems argument, not a prose recap.
The paper’s thesis lives or dies on hidden task structure. Toggle the same workload between a unified model and an explicit planner-router-specialist graph.
The optimistic theorem needs clean manifolds, identifiable task type, and reliable interfaces. Hover the matrix to see where the claimed advantage strengthens or collapses.
These are illustrative curves, not reported results. They encode the episode’s core tension: decomposition can shrink learning burden while routing, verification, and latency add new costs.
The position paper sits at the intersection of intelligence definitions, hierarchical planning, graph inductive bias, decomposed prompting, and recent agent workflow research.