<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="stoisser-etal-2025-sparks">
    <titleInfo>
        <title>Sparks of Tabular Reasoning via Text2SQL Reinforcement Learning</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Josefa</namePart>
        <namePart type="given">Lia</namePart>
        <namePart type="family">Stoisser</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Marc</namePart>
        <namePart type="given">Boubnovski</namePart>
        <namePart type="family">Martell</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Julien</namePart>
        <namePart type="family">Fauqueur</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2025-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 4th Table Representation Learning Workshop</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Shuaichen</namePart>
            <namePart type="family">Chang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Madelon</namePart>
            <namePart type="family">Hulsebos</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Qian</namePart>
            <namePart type="family">Liu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Wenhu</namePart>
            <namePart type="family">Chen</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Huan</namePart>
            <namePart type="family">Sun</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Vienna, Austria</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-268-8</identifier>
    </relatedItem>
    <abstract>This work reframes the Text-to-SQL task as a pathway for teaching large language models (LLMs) to reason over and manipulate tabular data—moving beyond the traditional focus on query generation. We propose a two-stage framework that leverages SQL supervision to develop transferable table reasoning capabilities. First, we synthesize detailed chain-of-thought (CoT) traces from real-world SQL queries, providing step-by-step, clause-level supervision that teaches the model how to traverse, filter, and aggregate table fields. Second, we introduce a Group Relative Policy Optimization (GRPO) reinforcement learning objective that connects SQL execution accuracy to generalizable reasoning by encouraging steps that extend beyond task-specific syntax and transfer across datasets.Empirically, our approach improves performance on standard Text-to-SQL benchmarks and achieves substantial gains on reasoning-intensive datasets such as BIRD, CRT-QA and Tablebench, demonstrating enhanced generalization and interpretability. Specifically, the distilled-quantized LLaMA-8B model achieved a 34% relative increase in exact match scores on CRT-QA when trained on Text-to-SQL tasks, while Qwen-2.5-7B achieved a 10% and Qwen-2.5-14B a 6% relative increase. These results suggest that SQL can serve not only as a target formalism but also as an effective scaffold for learning robust, transferable reasoning over structured data.</abstract>
    <identifier type="citekey">stoisser-etal-2025-sparks</identifier>
    <identifier type="doi">10.18653/v1/2025.trl-1.20</identifier>
    <location>
        <url>https://aclanthology.org/2025.trl-1.20/</url>
    </location>
    <part>
        <date>2025-07</date>
        <extent unit="page">
            <start>229</start>
            <end>240</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Sparks of Tabular Reasoning via Text2SQL Reinforcement Learning
%A Stoisser, Josefa Lia
%A Martell, Marc Boubnovski
%A Fauqueur, Julien
%Y Chang, Shuaichen
%Y Hulsebos, Madelon
%Y Liu, Qian
%Y Chen, Wenhu
%Y Sun, Huan
%S Proceedings of the 4th Table Representation Learning Workshop
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-268-8
%F stoisser-etal-2025-sparks
%X This work reframes the Text-to-SQL task as a pathway for teaching large language models (LLMs) to reason over and manipulate tabular data—moving beyond the traditional focus on query generation. We propose a two-stage framework that leverages SQL supervision to develop transferable table reasoning capabilities. First, we synthesize detailed chain-of-thought (CoT) traces from real-world SQL queries, providing step-by-step, clause-level supervision that teaches the model how to traverse, filter, and aggregate table fields. Second, we introduce a Group Relative Policy Optimization (GRPO) reinforcement learning objective that connects SQL execution accuracy to generalizable reasoning by encouraging steps that extend beyond task-specific syntax and transfer across datasets.Empirically, our approach improves performance on standard Text-to-SQL benchmarks and achieves substantial gains on reasoning-intensive datasets such as BIRD, CRT-QA and Tablebench, demonstrating enhanced generalization and interpretability. Specifically, the distilled-quantized LLaMA-8B model achieved a 34% relative increase in exact match scores on CRT-QA when trained on Text-to-SQL tasks, while Qwen-2.5-7B achieved a 10% and Qwen-2.5-14B a 6% relative increase. These results suggest that SQL can serve not only as a target formalism but also as an effective scaffold for learning robust, transferable reasoning over structured data.
%R 10.18653/v1/2025.trl-1.20
%U https://aclanthology.org/2025.trl-1.20/
%U https://doi.org/10.18653/v1/2025.trl-1.20
%P 229-240

Markdown (Informal)

[Sparks of Tabular Reasoning via Text2SQL Reinforcement Learning](https://aclanthology.org/2025.trl-1.20/) (Stoisser et al., TRL 2025)

• Sparks of Tabular Reasoning via Text2SQL Reinforcement Learning (Stoisser et al., TRL 2025)

ACL

• Josefa Lia Stoisser, Marc Boubnovski Martell, and Julien Fauqueur. 2025. Sparks of Tabular Reasoning via Text2SQL Reinforcement Learning. In Proceedings of the 4th Table Representation Learning Workshop, pages 229–240, Vienna, Austria. Association for Computational Linguistics.