Systematic exploration of heterogeneous model assignments across ROMA roles

Investigate systematically the effects of assigning different foundation models to ROMA’s Atomizer, Planner, type-specific Executor, and Aggregator components, and determine how heterogeneous role–model allocations impact the framework’s performance across tasks.

Background

ROMA is designed to be model-agnostic and explicitly supports heterogeneous model assignment across its modular roles (Atomizer, Planner, Executors, Aggregator). However, in the reported experiments the authors primarily use a single base model per task to isolate the contribution of the framework’s execution structure, deferring a systematic study of cross-model interactions.

A thorough investigation of heterogeneous role–model pairing could clarify how to exploit model-specific strengths for decomposition, execution, and aggregation, and reveal trade-offs in accuracy, latency, and cost when mixing models within ROMA’s recursive control loop.