@inproceedings{liu-etal-2026-evidocvlm,
title = "{E}vi{D}oc{VLM}: Evidence-Aligned Reinforcement Learning for Multi-Page Document Question Answering",
author = "Liu, Anran and
Cao, Zongsheng and
Xie, Jun and
Chen, Feng and
Wang, Zigan",
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.1448/",
pages = "30865--30875",
ISBN = "979-8-89176-462-0",
abstract = "Models can answer questions from retrieved document pages while producing citations that are valid but irrelevant to the answer. This often happens when citation quality and answer correctness are optimized as separate signals, so the model can learn to cite text without relying on it. We introduce \textbf{EviDocVLM}, a training method for multi-page document question answering that links selected support to answer behavior. A structured source map provides citation addresses for page content. EviDocVLM first uses identifier validity and quotation fidelity to derive a continuous coefficient for the trajectory-level answer reward, thereby increasing traceable support without making every citation error a hard failure. It then applies Evidence-Aligned GRPO, which reuses full-context rollout log-probabilities as a frozen target for one training-only evidence-view pass. The resulting sampled-token alignment encourages cited spans to support the answer without an external grader, extra rollout, or inference-time module. Experiments on document parsing and multi-page document QA benchmarks show stronger answer performance and consistent gains from the evidence-conditioned objective under complex visual layouts."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-evidocvlm">
<titleInfo>
<title>EviDocVLM: Evidence-Aligned Reinforcement Learning for Multi-Page Document Question Answering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anran</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zongsheng</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Xie</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feng</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zigan</namePart>
<namePart type="family">Wang</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>Models can answer questions from retrieved document pages while producing citations that are valid but irrelevant to the answer. This often happens when citation quality and answer correctness are optimized as separate signals, so the model can learn to cite text without relying on it. We introduce EviDocVLM, a training method for multi-page document question answering that links selected support to answer behavior. A structured source map provides citation addresses for page content. EviDocVLM first uses identifier validity and quotation fidelity to derive a continuous coefficient for the trajectory-level answer reward, thereby increasing traceable support without making every citation error a hard failure. It then applies Evidence-Aligned GRPO, which reuses full-context rollout log-probabilities as a frozen target for one training-only evidence-view pass. The resulting sampled-token alignment encourages cited spans to support the answer without an external grader, extra rollout, or inference-time module. Experiments on document parsing and multi-page document QA benchmarks show stronger answer performance and consistent gains from the evidence-conditioned objective under complex visual layouts.</abstract>
<identifier type="citekey">liu-etal-2026-evidocvlm</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1448/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>30865</start>
<end>30875</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T EviDocVLM: Evidence-Aligned Reinforcement Learning for Multi-Page Document Question Answering
%A Liu, Anran
%A Cao, Zongsheng
%A Xie, Jun
%A Chen, Feng
%A Wang, Zigan
%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 liu-etal-2026-evidocvlm
%X Models can answer questions from retrieved document pages while producing citations that are valid but irrelevant to the answer. This often happens when citation quality and answer correctness are optimized as separate signals, so the model can learn to cite text without relying on it. We introduce EviDocVLM, a training method for multi-page document question answering that links selected support to answer behavior. A structured source map provides citation addresses for page content. EviDocVLM first uses identifier validity and quotation fidelity to derive a continuous coefficient for the trajectory-level answer reward, thereby increasing traceable support without making every citation error a hard failure. It then applies Evidence-Aligned GRPO, which reuses full-context rollout log-probabilities as a frozen target for one training-only evidence-view pass. The resulting sampled-token alignment encourages cited spans to support the answer without an external grader, extra rollout, or inference-time module. Experiments on document parsing and multi-page document QA benchmarks show stronger answer performance and consistent gains from the evidence-conditioned objective under complex visual layouts.
%U https://aclanthology.org/2026.findings-emnlp.1448/
%P 30865-30875
Markdown (Informal)
[EviDocVLM: Evidence-Aligned Reinforcement Learning for Multi-Page Document Question Answering](https://aclanthology.org/2026.findings-emnlp.1448/) (Liu et al., Findings 2026)
ACL