@inproceedings{younis-2026-ahmed,
title = "Ahmed Younis at {A}r{G}uard Shared Tasks: Calibration Is Most of What You Tune: Decode-Time Structure and Threshold Transport for {A}rabic Harmful-Content Detection",
author = "Younis, Ahmed",
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.61/",
pages = "465--469",
ISBN = "979-8-89176-497-2",
abstract = "We describe our systems for all four ArGuard 2026 subtasks: binary and fine-grained Arabic hateful-meme detection (A1/A2) and binary and granular harmful-prompt detection (B1/B2). Our systems are deliberately conventional (a LoRA-tuned Qwen3-VL-8B for memes, a nine-encoder ensemble for prompts) because our central finding is that most of the achievable headroom lay not in the model but in decode-time calibration: per-class thresholds refit under structural constraints, rank-based decoding against known or published class priors, and threshold transport under distribution shift. On subtask A2, our submissions rose from 0.248 to 0.419 macro-F1; roughly 0.15 of that gain came from decode-time changes against under 0.03 from model-side interventions, and a fixed-model ablation on the ensemble{'}s char-ngram member attributes +0.070 to the decode stack alone. On B1, recalibrating a single threshold recovered +0.16 macro-F1 after we detected that the blind test was rebalanced to approximately 49{\%} unsafe against an 80{\%} unsafe training prior. Our final systems ranked 1st on subtask A2 (macro-F1 0.4194; 0.3913 without per-class scoring feedback), 3rd on A1 (0.7889), 3rd on B2 (0.7695), and 7th on B1 (0.7552). We additionally report a benchmark audit conducted during participation: a sequential-id label leak in the Task B development release (macro-F1 1.000 from ids alone), identified before the test release, left unexploited, and absent from the re-indexed test files; latent taxonomy structure that contradicts the task description; and a quantified gap between development and test distributions."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="younis-2026-ahmed">
<titleInfo>
<title>Ahmed Younis at ArGuard Shared Tasks: Calibration Is Most of What You Tune: Decode-Time Structure and Threshold Transport for Arabic Harmful-Content Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Younis</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 our systems for all four ArGuard 2026 subtasks: binary and fine-grained Arabic hateful-meme detection (A1/A2) and binary and granular harmful-prompt detection (B1/B2). Our systems are deliberately conventional (a LoRA-tuned Qwen3-VL-8B for memes, a nine-encoder ensemble for prompts) because our central finding is that most of the achievable headroom lay not in the model but in decode-time calibration: per-class thresholds refit under structural constraints, rank-based decoding against known or published class priors, and threshold transport under distribution shift. On subtask A2, our submissions rose from 0.248 to 0.419 macro-F1; roughly 0.15 of that gain came from decode-time changes against under 0.03 from model-side interventions, and a fixed-model ablation on the ensemble’s char-ngram member attributes +0.070 to the decode stack alone. On B1, recalibrating a single threshold recovered +0.16 macro-F1 after we detected that the blind test was rebalanced to approximately 49% unsafe against an 80% unsafe training prior. Our final systems ranked 1st on subtask A2 (macro-F1 0.4194; 0.3913 without per-class scoring feedback), 3rd on A1 (0.7889), 3rd on B2 (0.7695), and 7th on B1 (0.7552). We additionally report a benchmark audit conducted during participation: a sequential-id label leak in the Task B development release (macro-F1 1.000 from ids alone), identified before the test release, left unexploited, and absent from the re-indexed test files; latent taxonomy structure that contradicts the task description; and a quantified gap between development and test distributions.</abstract>
<identifier type="citekey">younis-2026-ahmed</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-sharedtasks.61/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>465</start>
<end>469</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings %T Ahmed Younis at ArGuard Shared Tasks: Calibration Is Most of What You Tune: Decode-Time Structure and Threshold Transport for Arabic Harmful-Content Detection %A Younis, Ahmed %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 younis-2026-ahmed %X We describe our systems for all four ArGuard 2026 subtasks: binary and fine-grained Arabic hateful-meme detection (A1/A2) and binary and granular harmful-prompt detection (B1/B2). Our systems are deliberately conventional (a LoRA-tuned Qwen3-VL-8B for memes, a nine-encoder ensemble for prompts) because our central finding is that most of the achievable headroom lay not in the model but in decode-time calibration: per-class thresholds refit under structural constraints, rank-based decoding against known or published class priors, and threshold transport under distribution shift. On subtask A2, our submissions rose from 0.248 to 0.419 macro-F1; roughly 0.15 of that gain came from decode-time changes against under 0.03 from model-side interventions, and a fixed-model ablation on the ensemble’s char-ngram member attributes +0.070 to the decode stack alone. On B1, recalibrating a single threshold recovered +0.16 macro-F1 after we detected that the blind test was rebalanced to approximately 49% unsafe against an 80% unsafe training prior. Our final systems ranked 1st on subtask A2 (macro-F1 0.4194; 0.3913 without per-class scoring feedback), 3rd on A1 (0.7889), 3rd on B2 (0.7695), and 7th on B1 (0.7552). We additionally report a benchmark audit conducted during participation: a sequential-id label leak in the Task B development release (macro-F1 1.000 from ids alone), identified before the test release, left unexploited, and absent from the re-indexed test files; latent taxonomy structure that contradicts the task description; and a quantified gap between development and test distributions. %U https://aclanthology.org/2026.arabicnlp-sharedtasks.61/ %P 465-469
Markdown (Informal)
[Ahmed Younis at ArGuard Shared Tasks: Calibration Is Most of What You Tune: Decode-Time Structure and Threshold Transport for Arabic Harmful-Content Detection](https://aclanthology.org/2026.arabicnlp-sharedtasks.61/) (Younis, ArabicNLP 2026)