@inproceedings{yamin-etal-2026-dynamos,
title = "Dynamos at {I}mage{E}val 2026 Shared Tasks: Metric-Aligned Selection and Cross-Family Ensembling for Vision-Language Hallucination Detection in {E}nglish and {A}rabic",
author = "Yamin, Muhammad and
Wahid, Sheikh Abdul and
Arshad, Ali",
editor = "Abdul-Mageed, Muhammad and
Elsayed, Tamer and
Bouamor, Houda and
El-Beltagy, Samhaa and
Al-Khalifa, Hend and
Abdelali, Ahmed and
Elmadany, AbdelRahim and
El Mekki, Abdellah and
Darwish, Kareem and
Alkhereyf, Sakhar and
Luqman, Hamzah and
Khalifa, Salam and
Magdy, Samar and
Alyafeai, Zaid and
Hamed, Injy and
Zaghouani, Wajdi and
Zitouni, Imed",
booktitle = "Proceedings of The Fourth {A}rabic Natural Language Processing Conference: Shared Tasks",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.arabicnlp-sharedtasks.85/",
pages = "681--688",
ISBN = "979-8-89176-497-2",
abstract = "We describe the Dynamos submission to ImageEval 2026 Task 1b, image-grounded hallucination detection: given an image and three statements, exactly one of which is grounded, a system must label each true or false. The official ranking metric, Contrastive Instability (CI, lower is better), reduces exactly to one minus the accuracy of picking the grounded statement once a system always outputs a single true label, which removes an entire failure mode (marking zero or several statements true) at no cost. Building on this reduction, we score each statement with an open vision-language model and always output the highest-scoring one as true. Starting from zero-shot Qwen2.5-VL, we low-rank fine-tune the model with a selection objective, train multiple random-seed adapters, and average their scores. Our main finding is that the strongest single model is not the strongest ensemble member: a 32B adapter trained on more data performs substantially better on the labelled devtest split but generalises worse than smaller 7B adapters to the blind test set, yet combining the two model families improves over either pure-family ensemble on both languages, despite the 32B family generalising worse on its own. Our final submitted systems reach CI 0.033 in English (2nd of 8 teams on the official leaderboard) and 0.056 in Modern Standard Arabic (3rd of 6 teams), both well below our zero-shot baselines of 0.058 (English) and 0.110 (MSA)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yamin-etal-2026-dynamos">
<titleInfo>
<title>Dynamos at ImageEval 2026 Shared Tasks: Metric-Aligned Selection and Cross-Family Ensembling for Vision-Language Hallucination Detection in English and Arabic</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Yamin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sheikh</namePart>
<namePart type="given">Abdul</namePart>
<namePart type="family">Wahid</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ali</namePart>
<namePart type="family">Arshad</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>Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Abdul-Mageed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tamer</namePart>
<namePart type="family">Elsayed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Houda</namePart>
<namePart type="family">Bouamor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samhaa</namePart>
<namePart type="family">El-Beltagy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hend</namePart>
<namePart type="family">Al-Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Abdelali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">AbdelRahim</namePart>
<namePart type="family">Elmadany</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Abdellah</namePart>
<namePart type="family">El Mekki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kareem</namePart>
<namePart type="family">Darwish</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakhar</namePart>
<namePart type="family">Alkhereyf</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hamzah</namePart>
<namePart type="family">Luqman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Salam</namePart>
<namePart type="family">Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samar</namePart>
<namePart type="family">Magdy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zaid</namePart>
<namePart type="family">Alyafeai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Injy</namePart>
<namePart type="family">Hamed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wajdi</namePart>
<namePart type="family">Zaghouani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Imed</namePart>
<namePart type="family">Zitouni</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">979-8-89176-497-2</identifier>
</relatedItem>
<abstract>We describe the Dynamos submission to ImageEval 2026 Task 1b, image-grounded hallucination detection: given an image and three statements, exactly one of which is grounded, a system must label each true or false. The official ranking metric, Contrastive Instability (CI, lower is better), reduces exactly to one minus the accuracy of picking the grounded statement once a system always outputs a single true label, which removes an entire failure mode (marking zero or several statements true) at no cost. Building on this reduction, we score each statement with an open vision-language model and always output the highest-scoring one as true. Starting from zero-shot Qwen2.5-VL, we low-rank fine-tune the model with a selection objective, train multiple random-seed adapters, and average their scores. Our main finding is that the strongest single model is not the strongest ensemble member: a 32B adapter trained on more data performs substantially better on the labelled devtest split but generalises worse than smaller 7B adapters to the blind test set, yet combining the two model families improves over either pure-family ensemble on both languages, despite the 32B family generalising worse on its own. Our final submitted systems reach CI 0.033 in English (2nd of 8 teams on the official leaderboard) and 0.056 in Modern Standard Arabic (3rd of 6 teams), both well below our zero-shot baselines of 0.058 (English) and 0.110 (MSA).</abstract>
<identifier type="citekey">yamin-etal-2026-dynamos</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-sharedtasks.85/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>681</start>
<end>688</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Dynamos at ImageEval 2026 Shared Tasks: Metric-Aligned Selection and Cross-Family Ensembling for Vision-Language Hallucination Detection in English and Arabic
%A Yamin, Muhammad
%A Wahid, Sheikh Abdul
%A Arshad, Ali
%Y Abdul-Mageed, Muhammad
%Y Elsayed, Tamer
%Y Bouamor, Houda
%Y El-Beltagy, Samhaa
%Y Al-Khalifa, Hend
%Y Abdelali, Ahmed
%Y Elmadany, AbdelRahim
%Y El Mekki, Abdellah
%Y Darwish, Kareem
%Y Alkhereyf, Sakhar
%Y Luqman, Hamzah
%Y Khalifa, Salam
%Y Magdy, Samar
%Y Alyafeai, Zaid
%Y Hamed, Injy
%Y Zaghouani, Wajdi
%Y Zitouni, Imed
%S Proceedings of The Fourth Arabic Natural Language Processing Conference: Shared Tasks
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-497-2
%F yamin-etal-2026-dynamos
%X We describe the Dynamos submission to ImageEval 2026 Task 1b, image-grounded hallucination detection: given an image and three statements, exactly one of which is grounded, a system must label each true or false. The official ranking metric, Contrastive Instability (CI, lower is better), reduces exactly to one minus the accuracy of picking the grounded statement once a system always outputs a single true label, which removes an entire failure mode (marking zero or several statements true) at no cost. Building on this reduction, we score each statement with an open vision-language model and always output the highest-scoring one as true. Starting from zero-shot Qwen2.5-VL, we low-rank fine-tune the model with a selection objective, train multiple random-seed adapters, and average their scores. Our main finding is that the strongest single model is not the strongest ensemble member: a 32B adapter trained on more data performs substantially better on the labelled devtest split but generalises worse than smaller 7B adapters to the blind test set, yet combining the two model families improves over either pure-family ensemble on both languages, despite the 32B family generalising worse on its own. Our final submitted systems reach CI 0.033 in English (2nd of 8 teams on the official leaderboard) and 0.056 in Modern Standard Arabic (3rd of 6 teams), both well below our zero-shot baselines of 0.058 (English) and 0.110 (MSA).
%U https://aclanthology.org/2026.arabicnlp-sharedtasks.85/
%P 681-688
Markdown (Informal)
[Dynamos at ImageEval 2026 Shared Tasks: Metric-Aligned Selection and Cross-Family Ensembling for Vision-Language Hallucination Detection in English and Arabic](https://aclanthology.org/2026.arabicnlp-sharedtasks.85/) (Yamin et al., ArabicNLP 2026)
ACL