arXiv meta-router beats static workflows with 100% success at 43% lower cost

A new arXiv paper from Natan Vidra, Alina Kapanova, Arun Kanhai, and Spurthi Setty tackles a core challenge in agentic AI: deciding which operations to run before producing an answer. Their system, a compositional meta-router, learns to compose heterogeneous actions—direct answering, decomposition, retrieval, code execution, delegation, and verification—directly from raw task text, rather than merely picking an LLM endpoint or retrieval depth.

The authors built an executable benchmark spanning data analysis, frozen-corpus research, and document processing, with 216 training, 72 development, 108 held-out test, and 108 locked challenge tasks. Outcomes are machine-checked after execution, enabling precise evaluation. The meta-router uses independent regularized logistic heads to predict operation probabilities from word and character features, then greedily composes operations under route-cost and action-count budgets. On the held-out test, it achieves 100% success versus 93.5% for strong static workflows, with 43% lower cost, and far outperforms a learned one-shot router (56.5%). On the untouched challenge split with shifted vocabulary, success drops to 75.9%, trailing static routing but remaining 49% cheaper and exceeding one-shot routing by 34.3 points. The authors identify lexical generalization, not route execution, as the principal limitation, noting the work is a reproducible testbed and bounded proof of concept rather than evidence of live-LLM performance.