RA-MoWE is a framework that leverages workflow‑affinity embeddings to cluster queries and generate reusable expert workflows. Each embedding records how well a fixed set of reference workflows solves a query, exposing similarity in effective reasoning strategies. The system clusters queries, uses the cluster’s average embedding to initialize a specialized workflow, and refines it with execution feedback. An embedding encoder then predicts these embeddings directly from query text, allowing new queries to select an appropriate expert without first running the reference workflows. Experiments on a 300‑query test set drawn from four benchmarks covering mathematics, science, and programming show that RA-MoWE improves average task score by 4.04 percentage points over simple reference‑workflow selection while cutting language‑model calls by 27.7% at inference.
Review