@inproceedings{ma-etal-2026-team,
title = "Team Falcons at {I}mage{E}val 2026 Shared Tasks:Structure-Constrained {QL}o{RA} Fine-Tuning and Dual-Model Consensus Verification",
author = "Ma, Pan and
Liujun and
Liu, Junlin",
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.88/",
pages = "703--710",
ISBN = "979-8-89176-497-2",
abstract = "We present the Team Falcons system for the ImageEval 2026 Task 1b English track. The task provides culturally grounded images from the OASIS/EverydayMMQA resource family and three English statements per image, exactly one of which is visually supported(CITATION). Rather than treating the statements as independent binary decisions, we exploit this structural constraint and formulate each item as image-conditioned three-way classification. In Stage 1, we fine-tune Qwen3-VL-8B(CITATION) with 4-bit QLoRA(CITATION) and estimate uncertainty from the first-token probabilities assigned to A, B, and C. Stage 2 we send only examples whose top-two probability margin is below 0.90 to Gemini 3.6 Flash and GPT-5.4-mini. Their answer replaces the Stage 1 prediction only when both verifiers return the same valid option. On the official test set, Stage 1 obtains a Contrastive Instability (CI) of 0.0360 and Combined Accuracy of 0.9640. Consensus verification improves these scores to 0.0290 and 0.9710, respectively, placing our submission first on the Task 1b English leaderboard. After test labels were released, we found that the verifiers agreed on 91 of 138 routed items and changed 17 predictions: 12 corrections and 5 harmful revisions. These results show that structure-aware fine-tuning and conservative uncertainty routing can effectively combine a local open-weight VLM with proprietary multimodal verifiers."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ma-etal-2026-team">
<titleInfo>
<title>Team Falcons at ImageEval 2026 Shared Tasks:Structure-Constrained QLoRA Fine-Tuning and Dual-Model Consensus Verification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pan</namePart>
<namePart type="family">Ma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name>
<namePart>Liujun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junlin</namePart>
<namePart type="family">Liu</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 present the Team Falcons system for the ImageEval 2026 Task 1b English track. The task provides culturally grounded images from the OASIS/EverydayMMQA resource family and three English statements per image, exactly one of which is visually supported(CITATION). Rather than treating the statements as independent binary decisions, we exploit this structural constraint and formulate each item as image-conditioned three-way classification. In Stage 1, we fine-tune Qwen3-VL-8B(CITATION) with 4-bit QLoRA(CITATION) and estimate uncertainty from the first-token probabilities assigned to A, B, and C. Stage 2 we send only examples whose top-two probability margin is below 0.90 to Gemini 3.6 Flash and GPT-5.4-mini. Their answer replaces the Stage 1 prediction only when both verifiers return the same valid option. On the official test set, Stage 1 obtains a Contrastive Instability (CI) of 0.0360 and Combined Accuracy of 0.9640. Consensus verification improves these scores to 0.0290 and 0.9710, respectively, placing our submission first on the Task 1b English leaderboard. After test labels were released, we found that the verifiers agreed on 91 of 138 routed items and changed 17 predictions: 12 corrections and 5 harmful revisions. These results show that structure-aware fine-tuning and conservative uncertainty routing can effectively combine a local open-weight VLM with proprietary multimodal verifiers.</abstract>
<identifier type="citekey">ma-etal-2026-team</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-sharedtasks.88/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>703</start>
<end>710</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Team Falcons at ImageEval 2026 Shared Tasks:Structure-Constrained QLoRA Fine-Tuning and Dual-Model Consensus Verification
%A Ma, Pan
%A Liu, Junlin
%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
%A Liujun
%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 ma-etal-2026-team
%X We present the Team Falcons system for the ImageEval 2026 Task 1b English track. The task provides culturally grounded images from the OASIS/EverydayMMQA resource family and three English statements per image, exactly one of which is visually supported(CITATION). Rather than treating the statements as independent binary decisions, we exploit this structural constraint and formulate each item as image-conditioned three-way classification. In Stage 1, we fine-tune Qwen3-VL-8B(CITATION) with 4-bit QLoRA(CITATION) and estimate uncertainty from the first-token probabilities assigned to A, B, and C. Stage 2 we send only examples whose top-two probability margin is below 0.90 to Gemini 3.6 Flash and GPT-5.4-mini. Their answer replaces the Stage 1 prediction only when both verifiers return the same valid option. On the official test set, Stage 1 obtains a Contrastive Instability (CI) of 0.0360 and Combined Accuracy of 0.9640. Consensus verification improves these scores to 0.0290 and 0.9710, respectively, placing our submission first on the Task 1b English leaderboard. After test labels were released, we found that the verifiers agreed on 91 of 138 routed items and changed 17 predictions: 12 corrections and 5 harmful revisions. These results show that structure-aware fine-tuning and conservative uncertainty routing can effectively combine a local open-weight VLM with proprietary multimodal verifiers.
%U https://aclanthology.org/2026.arabicnlp-sharedtasks.88/
%P 703-710
Markdown (Informal)
[Team Falcons at ImageEval 2026 Shared Tasks:Structure-Constrained QLoRA Fine-Tuning and Dual-Model Consensus Verification](https://aclanthology.org/2026.arabicnlp-sharedtasks.88/) (Ma et al., ArabicNLP 2026)
ACL