Spending Budget on a Better Pipeline Over a Mid-Tier Model Usually Beats Upgrading to a Frontier Model, Across 17 Text-to-SQL Configurations
arXiv 2608.28432 (2026-08-28, cs.AI/cs.CL/cs.DB) instantiates 17 paradigm-level configurations across five recurring modules of the in-context-learning text-to-SQL pipeline under one controlled implementation, attributing each module's marginal accuracy contribution and its token cost across four backbones. Execution-feedback refinement is the only paradigm whose benefit holds universally at consistently low cost; most other modules help only under backbone-dependent conditions. Token accounting shows input demand tracks pipeline structure while output demand tracks backbone generation behavior, and the headline guidance is that a fixed budget is often better spent on a more elaborate pipeline over a mid-tier backbone than on a frontier model with a lean pipeline, a tiered guideline that transferred to five additional backbones without per-paradigm search.
↳ Follow the thread