@inproceedings{he-etal-2026-local,
title = "Local Evidence, Global Credit: Visual Evidence-Calibrated Policy Optimization for Multimodal Reasoning",
author = "He, Zijing and
Deng, Zhongyi and
Lan, Kankan and
Chen, C.L.Philip and
Chen, Zeyu and
Gong, Enlei and
Zhang, Tong",
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.1818/",
doi = "10.18653/2026.findings-emnlp.1818",
pages = "38844--38858",
ISBN = "979-8-89176-462-0",
abstract = "Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning capability of large vision-language models (LVLMs). Most RLVR methods for LVLMs follow an outcome-driven paradigm, using the correctness of final answer as supervision. However, in multimodal reasoning, this supervision overlooks a critical question: whether a successful reasoning trajectory is actually grounded in visual evidence. To address this limitation, we propose Visual Evidence-Calibrated Policy Optimization (VEPO), which turns visual grounding into an explicit trajectory-level credit calibration signal. VEPO first measures token-level visual responses by comparing the model{'}s likelihood under the original and perturbed visual inputs. It then estimates a visual background for the rollout-group and retains responses above the background as sparse visual anchors into a rollout-level evidence score. This score calibrates outcome-driven optimization and guides the policy toward successful trajectories with stronger visual support. Experiments on six multimodal reasoning benchmarks show that VEPO achieves strong overall performance against RLVR baselines without requiring additional annotations, demonstrating its scalability for multimodal policy optimization."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="he-etal-2026-local">
<titleInfo>
<title>Local Evidence, Global Credit: Visual Evidence-Calibrated Policy Optimization for Multimodal Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zijing</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhongyi</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kankan</namePart>
<namePart type="family">Lan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">C.L.Philip</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeyu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Enlei</namePart>
<namePart type="family">Gong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</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>Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning capability of large vision-language models (LVLMs). Most RLVR methods for LVLMs follow an outcome-driven paradigm, using the correctness of final answer as supervision. However, in multimodal reasoning, this supervision overlooks a critical question: whether a successful reasoning trajectory is actually grounded in visual evidence. To address this limitation, we propose Visual Evidence-Calibrated Policy Optimization (VEPO), which turns visual grounding into an explicit trajectory-level credit calibration signal. VEPO first measures token-level visual responses by comparing the model’s likelihood under the original and perturbed visual inputs. It then estimates a visual background for the rollout-group and retains responses above the background as sparse visual anchors into a rollout-level evidence score. This score calibrates outcome-driven optimization and guides the policy toward successful trajectories with stronger visual support. Experiments on six multimodal reasoning benchmarks show that VEPO achieves strong overall performance against RLVR baselines without requiring additional annotations, demonstrating its scalability for multimodal policy optimization.</abstract>
<identifier type="citekey">he-etal-2026-local</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1818</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1818/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38844</start>
<end>38858</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Local Evidence, Global Credit: Visual Evidence-Calibrated Policy Optimization for Multimodal Reasoning
%A He, Zijing
%A Deng, Zhongyi
%A Lan, Kankan
%A Chen, C.L.Philip
%A Chen, Zeyu
%A Gong, Enlei
%A Zhang, Tong
%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 he-etal-2026-local
%X Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning capability of large vision-language models (LVLMs). Most RLVR methods for LVLMs follow an outcome-driven paradigm, using the correctness of final answer as supervision. However, in multimodal reasoning, this supervision overlooks a critical question: whether a successful reasoning trajectory is actually grounded in visual evidence. To address this limitation, we propose Visual Evidence-Calibrated Policy Optimization (VEPO), which turns visual grounding into an explicit trajectory-level credit calibration signal. VEPO first measures token-level visual responses by comparing the model’s likelihood under the original and perturbed visual inputs. It then estimates a visual background for the rollout-group and retains responses above the background as sparse visual anchors into a rollout-level evidence score. This score calibrates outcome-driven optimization and guides the policy toward successful trajectories with stronger visual support. Experiments on six multimodal reasoning benchmarks show that VEPO achieves strong overall performance against RLVR baselines without requiring additional annotations, demonstrating its scalability for multimodal policy optimization.
%R 10.18653/2026.findings-emnlp.1818
%U https://aclanthology.org/2026.findings-emnlp.1818/
%U https://doi.org/10.18653/2026.findings-emnlp.1818
%P 38844-38858
Markdown (Informal)
[Local Evidence, Global Credit: Visual Evidence-Calibrated Policy Optimization for Multimodal Reasoning](https://aclanthology.org/2026.findings-emnlp.1818/) (He et al., Findings 2026)
ACL
- Zijing He, Zhongyi Deng, Kankan Lan, C.L.Philip Chen, Zeyu Chen, Enlei Gong, and Tong Zhang. 2026. Local Evidence, Global Credit: Visual Evidence-Calibrated Policy Optimization for Multimodal Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 38844–38858, Budapest, Hungary. Association for Computational Linguistics.