MOSAIC: MODULAR ORCHESTRATION OF SPECIALISED AI COMPONENTS

Abstract

Large frontier language models achieve strong general-purpose performance, but they are expensive to run. A natural question is whether a system of smaller, specialised components, combined with structured task decomposition, can match the accuracy of a frontier model on a representative set of tasks, while using frontier-scale inference only rarely. We propose MOSAIC, an architecture in which a query is resolved through a strict hierarchy of hardcoded orchestration logic, not through routing between pre-trained models of different sizes. Each level of the hierarchy receives a task specification, applies fixed control logic to decide whether and how to narrow it, and calls a specialised model directly once a sub-task is clearly scoped. We describe three mechanisms that stack together and are intended to reduce inference cost compared with a single dense model: model-scale substitution, early termination, and cumulative partial resolution. We state this as a single efficiency hypothesis (H1). We describe the architecture, state the hypothesis explicitly, formalise the structural guarantees of the hierarchy, and propose an evaluation plan to test the hypothesis directly. 1

Keywords

• hierarchical task decomposition
• multi-agent orchestration
• compute-efficient inference
• small language models
• bounded retry termination