@inproceedings{jiang-etal-2026-anchored,
title = "Anchored Reasoning on Tables: Bridging the Verifiability Gap in Multimodal Reinforcement Learning",
author = "Jiang, Yubo and
An, Yitong and
Xie, Fengying and
Jiang, Zhiguo and
Zhang, Haopeng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2306/",
pages = "49477--49497",
ISBN = "979-8-89176-462-0",
abstract = "The Verifiability Gap in table reasoning arises when multimodal LLMs produce correct answers while grounding intermediate steps in unsupported visual evidence. Intermediate visual claims are often not exposed in a symbolically addressable form, making standard RL with verifiable rewards unable to reliably inspect whether a reasoning step is actually supported by the image. We mitigate this gap for table reasoning with Anchored Reasoning, which turns each reasoning step into a checkable (coordinate, value, structural role) claim over the table grid. We instantiate this paradigm as V-tableR1: a policy VLM generates anchor-laden visual chains-of-thought, while a learned critic VLM estimates anchor fidelity against the source image. The policy is trained with PGPO, a GRPO-based process-guided formulation that integrates critic-gated trajectory rewards with stabilization mechanisms for long anchor-laden rollouts. V-tableR1 (4B) achieves the best open-source average accuracy across seven tabular reasoning benchmarks, obtains state-of-the-art results on several individual datasets, surpasses open-source models up to $18\times$ its size in aggregate, and improves over its SFT baseline by up to 10.66{\%} absolute accuracy."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jiang-etal-2026-anchored">
<titleInfo>
<title>Anchored Reasoning on Tables: Bridging the Verifiability Gap in Multimodal Reinforcement Learning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yubo</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yitong</namePart>
<namePart type="family">An</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fengying</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhiguo</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haopeng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>The Verifiability Gap in table reasoning arises when multimodal LLMs produce correct answers while grounding intermediate steps in unsupported visual evidence. Intermediate visual claims are often not exposed in a symbolically addressable form, making standard RL with verifiable rewards unable to reliably inspect whether a reasoning step is actually supported by the image. We mitigate this gap for table reasoning with Anchored Reasoning, which turns each reasoning step into a checkable (coordinate, value, structural role) claim over the table grid. We instantiate this paradigm as V-tableR1: a policy VLM generates anchor-laden visual chains-of-thought, while a learned critic VLM estimates anchor fidelity against the source image. The policy is trained with PGPO, a GRPO-based process-guided formulation that integrates critic-gated trajectory rewards with stabilization mechanisms for long anchor-laden rollouts. V-tableR1 (4B) achieves the best open-source average accuracy across seven tabular reasoning benchmarks, obtains state-of-the-art results on several individual datasets, surpasses open-source models up to 18\times its size in aggregate, and improves over its SFT baseline by up to 10.66% absolute accuracy.</abstract>
<identifier type="citekey">jiang-etal-2026-anchored</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2306/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>49477</start>
<end>49497</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Anchored Reasoning on Tables: Bridging the Verifiability Gap in Multimodal Reinforcement Learning
%A Jiang, Yubo
%A An, Yitong
%A Xie, Fengying
%A Jiang, Zhiguo
%A Zhang, Haopeng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F jiang-etal-2026-anchored
%X The Verifiability Gap in table reasoning arises when multimodal LLMs produce correct answers while grounding intermediate steps in unsupported visual evidence. Intermediate visual claims are often not exposed in a symbolically addressable form, making standard RL with verifiable rewards unable to reliably inspect whether a reasoning step is actually supported by the image. We mitigate this gap for table reasoning with Anchored Reasoning, which turns each reasoning step into a checkable (coordinate, value, structural role) claim over the table grid. We instantiate this paradigm as V-tableR1: a policy VLM generates anchor-laden visual chains-of-thought, while a learned critic VLM estimates anchor fidelity against the source image. The policy is trained with PGPO, a GRPO-based process-guided formulation that integrates critic-gated trajectory rewards with stabilization mechanisms for long anchor-laden rollouts. V-tableR1 (4B) achieves the best open-source average accuracy across seven tabular reasoning benchmarks, obtains state-of-the-art results on several individual datasets, surpasses open-source models up to 18\times its size in aggregate, and improves over its SFT baseline by up to 10.66% absolute accuracy.
%U https://aclanthology.org/2026.findings-emnlp.2306/
%P 49477-49497
Markdown (Informal)
[Anchored Reasoning on Tables: Bridging the Verifiability Gap in Multimodal Reinforcement Learning](https://aclanthology.org/2026.findings-emnlp.2306/) (Jiang et al., Findings 2026)
ACL