@inproceedings{liu-etal-2026-ovip,
title = "{OV}i{P}: Online Vision-Language Preference Learning with Hallucination-Targeted Negatives",
author = "Liu, Shujun and
Wang, Siyuan and
Li, Zejun and
Wang, Jianxiang and
Cheng, Zeng and
Huang, Xuanjing and
Wei, Zhongyu",
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.674/",
pages = "14142--14167",
ISBN = "979-8-89176-462-0",
abstract = "Large vision-language models (LVLMs) remain prone to hallucination. While recent visual Direct Preference Optimization (DPO) methods try to fix this by contrasting positive images with heuristic or offline negatives for the same correct response. We show that such negatives are mostly ``trivial'': the model already assigns near-zero probability to the correct response, so the resulting gradient is uninformative or even spuriously misdirected. We propose OViP, an on-policy framework that synthesizes negative images on the fly from the model{'}s own hallucinated responses via a diffusion prior, producing semantically misaligned yet still high-likelihood negatives. Across five hallucination and four general benchmarks, OViP achieves better hallucination mitigation on LLaVA-1.5 series and Qwen2.5-VL-3B, while being 2 times more compute-efficient than SIMA and incurring no degradation on general capabilities."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-ovip">
<titleInfo>
<title>OViP: Online Vision-Language Preference Learning with Hallucination-Targeted Negatives</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shujun</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyuan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zejun</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianxiang</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeng</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuanjing</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhongyu</namePart>
<namePart type="family">Wei</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>Large vision-language models (LVLMs) remain prone to hallucination. While recent visual Direct Preference Optimization (DPO) methods try to fix this by contrasting positive images with heuristic or offline negatives for the same correct response. We show that such negatives are mostly “trivial”: the model already assigns near-zero probability to the correct response, so the resulting gradient is uninformative or even spuriously misdirected. We propose OViP, an on-policy framework that synthesizes negative images on the fly from the model’s own hallucinated responses via a diffusion prior, producing semantically misaligned yet still high-likelihood negatives. Across five hallucination and four general benchmarks, OViP achieves better hallucination mitigation on LLaVA-1.5 series and Qwen2.5-VL-3B, while being 2 times more compute-efficient than SIMA and incurring no degradation on general capabilities.</abstract>
<identifier type="citekey">liu-etal-2026-ovip</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.674/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14142</start>
<end>14167</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T OViP: Online Vision-Language Preference Learning with Hallucination-Targeted Negatives
%A Liu, Shujun
%A Wang, Siyuan
%A Li, Zejun
%A Wang, Jianxiang
%A Cheng, Zeng
%A Huang, Xuanjing
%A Wei, Zhongyu
%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-ovip
%X Large vision-language models (LVLMs) remain prone to hallucination. While recent visual Direct Preference Optimization (DPO) methods try to fix this by contrasting positive images with heuristic or offline negatives for the same correct response. We show that such negatives are mostly “trivial”: the model already assigns near-zero probability to the correct response, so the resulting gradient is uninformative or even spuriously misdirected. We propose OViP, an on-policy framework that synthesizes negative images on the fly from the model’s own hallucinated responses via a diffusion prior, producing semantically misaligned yet still high-likelihood negatives. Across five hallucination and four general benchmarks, OViP achieves better hallucination mitigation on LLaVA-1.5 series and Qwen2.5-VL-3B, while being 2 times more compute-efficient than SIMA and incurring no degradation on general capabilities.
%U https://aclanthology.org/2026.findings-emnlp.674/
%P 14142-14167
Markdown (Informal)
[OViP: Online Vision-Language Preference Learning with Hallucination-Targeted Negatives](https://aclanthology.org/2026.findings-emnlp.674/) (Liu et al., Findings 2026)
ACL