Memon, Ibrahim, Karamzade, Armin, Jhaveri, M. Hanif, Fox, Roy · Nov 13, 2026
Ibrahim Memon and coauthors present LuthorSQL, a one-pass text-to-SQL model
Markdown · JSON · Original source
Ibrahim Memon, Armin Karamzade, M. Hanif Jhaveri and Roy Fox present LuthorSQL as a local, rapid System 1 gate for a slower agent: it handles queries when confident and escalates uncertain ones. The authors aim to offload hosted-model calls, which they associate with delay, cost and exposure of organizational queries, without compromising accuracy. LuthorSQL makes a single forward pass to predict slot scores over a bounded grammar for query plans, then compiles and runs candidate plans and uses a trained validator to choose an answer. The authors report 68.03% execution accuracy on BIRD's hidden at 0.84 seconds per question on a single . They report maintaining that accuracy in 9 GB with 8-bit weights, and transfer to five database systems within 3.1 points of native SQLite accuracy. In a pre-registered confidence cascade over 227 questions on databases published after Opus 5.5's training cutoff, LuthorSQL answered nearly half locally and cut cost by 33.3% at the same accuracy. Their contamination probing reports that Opus reproduces 48–92% of public gold queries verbatim without seeing the schema; tool-assisted accuracy held on renamed schemas and reached 95.6–97.4% on post-cutoff databases.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 13, 2026 (2026-11-13).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T08:03:47.601Z. First observed: 2026-11-16T05:01:10.823Z. Retrieved: 2026-11-16T08:02:46.426Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the paper's abstract, not the full paper; it does not provide the detailed evaluation methodology or results needed to assess the reported measurements. The confidence-cascade results are reported for 227 questions; the abstract does not establish how they generalize beyond that test.
Open question
How would the confidence cascade's local-query share, -cost reduction and accuracy compare on a larger set of databases published after Opus 5.5's training cutoff?
Source-stated
The authors propose LuthorSQL as a local System 1 gate that handles queries when confident and escalates uncertain ones to a slower LLM agent.
We present LuthorSQL, a non-autoregressive 7B text-to-SQL model that serves as a rapid System 1 gate for a slower LLM agent (System 2), escalating queries only when uncertain.
Source-stated
LuthorSQL uses a single forward pass to predict slot scores over a bounded grammar, then compiles and runs candidate plans and selects an answer with a trained validator.
LuthorSQL makes a single forward pass to predict slot scores over a bounded grammar for query plans, compiles and runs up to 32 candidate plans, and chooses an answer using a trained validator.
Source-stated
The authors report 68.03% execution accuracy on BIRD's hidden test set at 0.84 seconds per question on a single GPU.
On its own, LuthorSQL reaches 68.03% execution accuracy on BIRD's hidden test set at 0.84 s per question on a single GPU
Source-stated
The authors report that LuthorSQL maintains its accuracy in 9 GB with 8-bit quantized weights and transfers zero-shot to five database systems within 3.1 points of its native SQLite accuracy.
maintains its accuracy in 9 GB with 8-bit quantized weights, and transfers zero-shot to Snowflake, BigQuery, MySQL, PostgreSQL, and DuckDB within 3.1 points of its native SQLite accuracy.
Source-stated
In a pre-registered confidence cascade over 227 questions on post-cutoff databases, the authors report that LuthorSQL answers nearly half locally and reduces API cost by 33.3% at the same accuracy.
Combined with Claude Opus 5.5 in a pre-registered confidence cascade over 227 questions on databases published after Opus's training cutoff, LuthorSQL answers nearly half of all questions locally and cuts API cost by 33.3% at the same accuracy.
Source-stated
The authors report contamination probing found Opus reproducing public benchmark gold queries verbatim without seeing the schema, while tool-assisted accuracy held on renamed schemas and was 95.6–97.4% on post-cutoff databases.
contamination probing shows that Opus reproduces 48-92% of public benchmark gold queries verbatim without seeing the schema; however, its tool-assisted accuracy holds on renamed schemas and reaches 95.6-97.4% on post-cutoff databases.
Source and contributor context
Memon, Ibrahim, Karamzade, Armin, Jhaveri, M. Hanif, Fox, Roy · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Zero-shot
Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).
API
An application programming interface: a defined way for software to request data or actions from another system.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: f075cd83-36d5-4308-8c68-7e27bb441b81.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T08:03:47.601Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4a798ff8f0942759a1369a6d489b8699c516ebba6a1a46199dc8bcebd26cf1b7. Source text hash: d40b1d58619a4f40d3a6cf8ff4a68d7d32ed6774ae307e5fc8e6db003407b891.
Revision and source-check history
Version 1 · 2026-11-16T08:03:47.601Z
Ibrahim Memon and coauthors present LuthorSQL, a one-pass text-to-SQL model
Ibrahim Memon, Armin Karamzade, M. Hanif Jhaveri and Roy Fox present LuthorSQL as a local, rapid System 1 gate for a slower agent: it handles queries when confident and escalates uncertain ones. The authors aim to offload hosted-model calls, which they associate with delay, cost and exposure of organizational queries, without compromising accuracy. LuthorSQL makes a single forward pass to predict slot scores over a bounded grammar for query plans, then compiles and runs candidate plans and uses a trained validator to choose an answer. The authors report 68.03% execution accuracy on BIRD's hidden at 0.84 seconds per question on a single . They report maintaining that accuracy in 9 GB with 8-bit weights, and transfer to five database systems within 3.1 points of native SQLite accuracy. In a pre-registered confidence cascade over 227 questions on databases published after Opus 5.5's training cutoff, LuthorSQL answered nearly half locally and cut cost by 33.3% at the same accuracy. Their contamination probing reports that Opus reproduces 48–92% of public gold queries verbatim without seeing the schema; tool-assisted accuracy held on renamed schemas and reached 95.6–97.4% on post-cutoff databases.
Original source
cs.AI · Article
Source details
Original source date: Nov 13, 2026 (2026-11-13).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T08:03:47.601Z. First observed: 2026-11-16T05:01:10.823Z. Retrieved: 2026-11-16T08:02:46.426Z.
Limits
The evidence is the paper's abstract, not the full paper; it does not provide the detailed evaluation methodology or results needed to assess the reported measurements. The confidence-cascade results are reported for 227 questions; the abstract does not establish how they generalize beyond that test.
Open question
How would the confidence cascade's local-query share, -cost reduction and accuracy compare on a larger set of databases published after Opus 5.5's training cutoff?
Source-stated
The authors propose LuthorSQL as a local System 1 gate that handles queries when confident and escalates uncertain ones to a slower LLM agent.
We present LuthorSQL, a non-autoregressive 7B text-to-SQL model that serves as a rapid System 1 gate for a slower LLM agent (System 2), escalating queries only when uncertain.
Source-stated
LuthorSQL uses a single forward pass to predict slot scores over a bounded grammar, then compiles and runs candidate plans and selects an answer with a trained validator.
LuthorSQL makes a single forward pass to predict slot scores over a bounded grammar for query plans, compiles and runs up to 32 candidate plans, and chooses an answer using a trained validator.
Source-stated
The authors report 68.03% execution accuracy on BIRD's hidden test set at 0.84 seconds per question on a single GPU.
On its own, LuthorSQL reaches 68.03% execution accuracy on BIRD's hidden test set at 0.84 s per question on a single GPU
Source-stated
The authors report that LuthorSQL maintains its accuracy in 9 GB with 8-bit quantized weights and transfers zero-shot to five database systems within 3.1 points of its native SQLite accuracy.
maintains its accuracy in 9 GB with 8-bit quantized weights, and transfers zero-shot to Snowflake, BigQuery, MySQL, PostgreSQL, and DuckDB within 3.1 points of its native SQLite accuracy.
Source-stated
In a pre-registered confidence cascade over 227 questions on post-cutoff databases, the authors report that LuthorSQL answers nearly half locally and reduces API cost by 33.3% at the same accuracy.
Combined with Claude Opus 5.5 in a pre-registered confidence cascade over 227 questions on databases published after Opus's training cutoff, LuthorSQL answers nearly half of all questions locally and cuts API cost by 33.3% at the same accuracy.
Source-stated
The authors report contamination probing found Opus reproducing public benchmark gold queries verbatim without seeing the schema, while tool-assisted accuracy held on renamed schemas and was 95.6–97.4% on post-cutoff databases.
contamination probing shows that Opus reproduces 48-92% of public benchmark gold queries verbatim without seeing the schema; however, its tool-assisted accuracy holds on renamed schemas and reaches 95.6-97.4% on post-cutoff databases.
Source and contributor context
Memon, Ibrahim, Karamzade, Armin, Jhaveri, M. Hanif, Fox, Roy · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Held-out set
Data kept aside during training and used only to test how well a model works on examples it has not seen.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Zero-shot
Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).
API
An application programming interface: a defined way for software to request data or actions from another system.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: f075cd83-36d5-4308-8c68-7e27bb441b81.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T08:03:47.601Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4a798ff8f0942759a1369a6d489b8699c516ebba6a1a46199dc8bcebd26cf1b7. Source text hash: d40b1d58619a4f40d3a6cf8ff4a68d7d32ed6774ae307e5fc8e6db003407b891.
uncertain · 2026-11-16T08:03:47.601Z · anthropic/claude-opus-5.5
Publication: f075cd83-36d5-4308-8c68-7e27bb441b81
attribution — pass: The draft correctly names Ibrahim Memon, Armin Karamzade, M. Hanif Jhaveri and Roy Fox as authors. It frames the results as the authors' reports. One cascade sentence drops the explicit 'the authors report' framing. Adding it back would be a small improvement, but the meaning stays attributed.
Authors:Ibrahim Memon, Armin Karamzade, M. Hanif Jhaveri, Roy Fox
summary — uncertain: The draft is faithful but hard for a non-specialist to follow. Central terms are never explained in plain language: text-to-SQL (turning natural-language questions into database queries), System 1/System 2 (fast first pass vs slower deliberate agent), forward pass, bounded grammar, execution accuracy, BIRD, zero-shot, 8-bit quantized, contamination probing and gold queries. Smallest repair: add a plain gloss of text-to-SQL and the System 1/System 2 cascade, or tooltips for the others. The draft also never says why Opus reproducing gold queries matters, namely that benchmark answers may be memorized.
We present LuthorSQL, a non-autoregressive 7B text-to-SQL model that serves as a rapid System 1 gate for a slower LLM agent (System 2), escalating queries only when uncertain.
date — pass: The originalPublished value, 2026-11-13, matches the v1 submission date on arXiv. It is not the later feed date.
[Submitted on 13 Nov 2026]
support — pass: All the figures match the abstract: 68.03%, 0.84 s, 9 GB, 8-bit, 3.1 points, 227 questions, nearly half, 33.3%, 48-92% and 95.6-97.4%. The claim that the system runs locally is supported by the abstract's framing around local hardware. The question about generalizing beyond 227 questions follows from the evidence and is not leading.
To what extent can local hardware already owned by an organization offload this burden without compromising the agent's accuracy?
media — pass: The selected media is the verified arXiv abstract page of the source itself. This is the most direct way for a reader to reach the paper.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referenceA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceDoing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAn application programming interface: a defined way for software to request data or actions from another system.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivAn application programming interface: a defined way for software to request data or actions from another system.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeA language model with many learned numerical settings, used to process and generate language.
Definition referenceData kept aside during training and used only to test how well a model works on examples it has not seen.
Definition referenceA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceDoing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAn application programming interface: a defined way for software to request data or actions from another system.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivAn application programming interface: a defined way for software to request data or actions from another system.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore Claude