@inproceedings{gajjar-subramaniakuppusamy-2026-rsat,
title = "{RSAT}: Structured Attribution Makes Small Language Models Faithful Table Reasoners",
author = "Gajjar, Jugal and
Subramaniakuppusamy, Kamalasankari",
editor = "Gupta, Vivek and
Ding, Kaize and
Kokel, Harsha and
Zhao, Yue and
Agarwal, Amit and
Wang, Yu and
Glass, Michael and
Zhang, Yu and
Srinivas, Kavitha and
Chen, Xiusi and
Hassanzadeh, Oktie and
Zhu, Qi and
Chang, Shuaichen and
Luo, Yuan",
booktitle = "Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the {LLM} Era ({SURG}e{LLM} 2026)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.surgellm-1.7/",
pages = "119--131",
ISBN = "979-8-89176-406-4",
abstract = "When a language model answers a table question, users have no way to verify which cells informed which reasoning steps. We introduce RSAT, a method that trains small language models (SLMs, 1{--}8B) to produce step-by-step reasoning with cell-level citations grounded in table evidence. Phase 1 (SFT) teaches a structured JSON output format from verified reasoning traces. Phase 2 (GRPO) optimizes a composite reward centered on NLI-based faithfulness, alongside citation validity and parsimony. Across six models from two families{---}Qwen2.5 (1.5B/3B/7B) and Llama3 (1B/3B/8B){---}RSAT improves faithfulness 3.7$\times$ over SFT alone (0.224$\rightarrow$0.826), with near-perfect citation validity (0.992). Post-hoc attribution collapses below 13{\%} format success, confirming that attribution must be integrated into reasoning, not retrofitted. Ablations show the faithfulness reward is essential: removing it drops faithfulness from 0.97 to 0.03."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gajjar-subramaniakuppusamy-2026-rsat">
<titleInfo>
<title>RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jugal</namePart>
<namePart type="family">Gajjar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kamalasankari</namePart>
<namePart type="family">Subramaniakuppusamy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vivek</namePart>
<namePart type="family">Gupta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaize</namePart>
<namePart type="family">Ding</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Harsha</namePart>
<namePart type="family">Kokel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amit</namePart>
<namePart type="family">Agarwal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Glass</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kavitha</namePart>
<namePart type="family">Srinivas</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiusi</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Oktie</namePart>
<namePart type="family">Hassanzadeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qi</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<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">Yuan</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-406-4</identifier>
</relatedItem>
<abstract>When a language model answers a table question, users have no way to verify which cells informed which reasoning steps. We introduce RSAT, a method that trains small language models (SLMs, 1–8B) to produce step-by-step reasoning with cell-level citations grounded in table evidence. Phase 1 (SFT) teaches a structured JSON output format from verified reasoning traces. Phase 2 (GRPO) optimizes a composite reward centered on NLI-based faithfulness, alongside citation validity and parsimony. Across six models from two families—Qwen2.5 (1.5B/3B/7B) and Llama3 (1B/3B/8B)—RSAT improves faithfulness 3.7\times over SFT alone (0.224\rightarrow0.826), with near-perfect citation validity (0.992). Post-hoc attribution collapses below 13% format success, confirming that attribution must be integrated into reasoning, not retrofitted. Ablations show the faithfulness reward is essential: removing it drops faithfulness from 0.97 to 0.03.</abstract>
<identifier type="citekey">gajjar-subramaniakuppusamy-2026-rsat</identifier>
<location>
<url>https://aclanthology.org/2026.surgellm-1.7/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>119</start>
<end>131</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners
%A Gajjar, Jugal
%A Subramaniakuppusamy, Kamalasankari
%Y Gupta, Vivek
%Y Ding, Kaize
%Y Kokel, Harsha
%Y Zhao, Yue
%Y Agarwal, Amit
%Y Wang, Yu
%Y Glass, Michael
%Y Zhang, Yu
%Y Srinivas, Kavitha
%Y Chen, Xiusi
%Y Hassanzadeh, Oktie
%Y Zhu, Qi
%Y Chang, Shuaichen
%Y Luo, Yuan
%S Proceedings of the First Workshop on Structured Understanding, Retrieval, and Generation in the LLM Era (SURGeLLM 2026)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-406-4
%F gajjar-subramaniakuppusamy-2026-rsat
%X When a language model answers a table question, users have no way to verify which cells informed which reasoning steps. We introduce RSAT, a method that trains small language models (SLMs, 1–8B) to produce step-by-step reasoning with cell-level citations grounded in table evidence. Phase 1 (SFT) teaches a structured JSON output format from verified reasoning traces. Phase 2 (GRPO) optimizes a composite reward centered on NLI-based faithfulness, alongside citation validity and parsimony. Across six models from two families—Qwen2.5 (1.5B/3B/7B) and Llama3 (1B/3B/8B)—RSAT improves faithfulness 3.7\times over SFT alone (0.224\rightarrow0.826), with near-perfect citation validity (0.992). Post-hoc attribution collapses below 13% format success, confirming that attribution must be integrated into reasoning, not retrofitted. Ablations show the faithfulness reward is essential: removing it drops faithfulness from 0.97 to 0.03.
%U https://aclanthology.org/2026.surgellm-1.7/
%P 119-131
Markdown (Informal)
[RSAT: Structured Attribution Makes Small Language Models Faithful Table Reasoners](https://aclanthology.org/2026.surgellm-1.7/) (Gajjar & Subramaniakuppusamy, SURGeLLM 2026)
ACL