ai-routing · 2026-11-17 · Tier 1

LuthorSQL: a one-pass local text-to-SQL model as System 1 for a slower LLM agent (2611.17780)

LuthorSQL: a one-pass local text-to-SQL model as System 1 for a slower LLM agent (2611.17780)

Source: arXiv:2611.17780 · HuggingFace Daily Papers 2026-11-16 · raw: raw/huggingface/2026-11-16-luthorsql-a-one-pass-text-to-sql-model-as-system-1-for-a-slo.md

TL;DR

A Claude Opus 5.5 agent answering 10,000 database questions a day costs over $70,000 a year at list price and sends sensitive queries to an external API. LuthorSQL is a 7B non-autoregressive text-to-SQL model: instead of writing SQL token by token, one forward pass scores slots in a bounded grammar of query plans. It compiles and executes up to 32 candidate plans and a trained validator picks the answer. Alone it reaches 68.03% execution accuracy on BIRD's hidden test set at 0.84 s per question on one GPU, keeps that accuracy in 9 GB with 8-bit weights, and transfers zero-shot to Snowflake, BigQuery, MySQL, PostgreSQL and DuckDB within 3.1 points of its SQLite score. As a confidence cascade (answer locally when confident, escalate to Opus otherwise), pre-registered over 227 questions on databases published after Opus's training cutoff, it answers nearly half the questions locally and cuts API cost 33.3% at the same accuracy. A contamination probe finds Opus reproduces 48-92% of public benchmark gold queries verbatim without seeing the schema, though its tool-assisted accuracy holds on renamed schemas (95.6-97.4%).

Half the questions never leave the building
The router is the local model's own confidence.
flowchart LR
  Q["Question<br/><small>plus schema</small>"] --> L["LuthorSQL 7B<br/><small>one-pass plan scores</small>"]
  L --> X["32 plans<br/><small>compiled and executed</small>"]
  X --> V{"Validator<br/><small>confident?</small>"}
  V -->|yes| A["Local answer<br/><small>0.84 s, on-prem</small>"]
  V -->|no| O["Opus agent<br/><small>tools, slower, paid</small>"]
  O --> A2["Escalated answer<br/><small>same accuracy</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class Q input
  class L,X,O core
  class V loop
  class A,A2 exit
  linkStyle 3 stroke:#2f9e44,stroke-width:2px
Blue is input, purple is compute, amber is the routing decision, green is an answer.

How this relates to prior wiki pages

  • Confirms the decision-only-compute thread from the 11-16 Media Zone: Claude Code's advisor mode (consult the strong model only at checkpoints) and Perplexity's single-pass Decider. LuthorSQL is the domain-specific, measured version: a cheap one-pass model gates the expensive agent.
  • Echoes Satya Nadella (Pragmatic Engineer, 11-16), who described Microsoft building "hybrid intelligence": a router choosing between unmetered local models and paid cloud models.
  • The contamination finding is a caution for every router trained or evaluated on public text-to-SQL benchmarks: the strong model's "ground truth" advantage may be memorized.
  • Concept page: LLM routing.

Gaps

  • 227 cascade questions is small; the 33.3% saving has wide uncertainty.
  • The grammar is bounded; queries outside it must always escalate, and the paper's share of those is the cap on local coverage.