@inproceedings{joshi-2026-yalla,
title = "Yalla at {AISA}-{A}rabic{FC}: Argument-Focused Fine-Tuning and Per-Argument Ensembling for {A}rabic Function Calling",
author = "Joshi, Manish",
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.208/",
pages = "1608--1613",
ISBN = "979-8-89176-497-2",
abstract = "We describe the Team Yalla submission to the AISA-ArabicFC shared task on Arabic function calling, which ranked second in both tracks (0.9273 Track A, 0.9395 Track B), within 0.0006 of the winning system. The headline metric, exact argument match (ArgEM), is unforgiving: a single wrong surface form loses the whole example. The provided targets bury the scored object inside a rationale: the tool name and argument dictionary are only 30.3{\%} of the target characters. We trained on the 30.3{\%}, which is our largest single gain (+0.020 ArgEM in a matched comparison). A matched follow-up shows that reasoning need not be removed entirely: emitting it after the arguments trails the JSON-only target by only 0.008 ArgEM. We add a reinforcement-learning stage that uses the official scorer as the reward, and combine three deliberately diverse fine-tunes with a per-argument majority vote, which repairs individual argument slots rather than choosing whole predictions. Every member is a LoRA fine-tune of a public 4B or 7B Qwen model trained on the official split alone, with no distillation and no external data. We note that two of the three voting members emit no rationale, so our Track B think field is attached after prediction from a separate reasoning-trained model whose predicted call for that row is identical: it is a post-hoc explanation, not the computation that produced the arguments. We report the experiments that failed alongside those that worked."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="joshi-2026-yalla">
<titleInfo>
<title>Yalla at AISA-ArabicFC: Argument-Focused Fine-Tuning and Per-Argument Ensembling for Arabic Function Calling</title>
</titleInfo>
<name type="personal">
<namePart type="given">Manish</namePart>
<namePart type="family">Joshi</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 Team Yalla submission to the AISA-ArabicFC shared task on Arabic function calling, which ranked second in both tracks (0.9273 Track A, 0.9395 Track B), within 0.0006 of the winning system. The headline metric, exact argument match (ArgEM), is unforgiving: a single wrong surface form loses the whole example. The provided targets bury the scored object inside a rationale: the tool name and argument dictionary are only 30.3% of the target characters. We trained on the 30.3%, which is our largest single gain (+0.020 ArgEM in a matched comparison). A matched follow-up shows that reasoning need not be removed entirely: emitting it after the arguments trails the JSON-only target by only 0.008 ArgEM. We add a reinforcement-learning stage that uses the official scorer as the reward, and combine three deliberately diverse fine-tunes with a per-argument majority vote, which repairs individual argument slots rather than choosing whole predictions. Every member is a LoRA fine-tune of a public 4B or 7B Qwen model trained on the official split alone, with no distillation and no external data. We note that two of the three voting members emit no rationale, so our Track B think field is attached after prediction from a separate reasoning-trained model whose predicted call for that row is identical: it is a post-hoc explanation, not the computation that produced the arguments. We report the experiments that failed alongside those that worked.</abstract>
<identifier type="citekey">joshi-2026-yalla</identifier>
<location>
<url>https://aclanthology.org/2026.arabicnlp-sharedtasks.208/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1608</start>
<end>1613</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Yalla at AISA-ArabicFC: Argument-Focused Fine-Tuning and Per-Argument Ensembling for Arabic Function Calling
%A Joshi, Manish
%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 joshi-2026-yalla
%X We describe the Team Yalla submission to the AISA-ArabicFC shared task on Arabic function calling, which ranked second in both tracks (0.9273 Track A, 0.9395 Track B), within 0.0006 of the winning system. The headline metric, exact argument match (ArgEM), is unforgiving: a single wrong surface form loses the whole example. The provided targets bury the scored object inside a rationale: the tool name and argument dictionary are only 30.3% of the target characters. We trained on the 30.3%, which is our largest single gain (+0.020 ArgEM in a matched comparison). A matched follow-up shows that reasoning need not be removed entirely: emitting it after the arguments trails the JSON-only target by only 0.008 ArgEM. We add a reinforcement-learning stage that uses the official scorer as the reward, and combine three deliberately diverse fine-tunes with a per-argument majority vote, which repairs individual argument slots rather than choosing whole predictions. Every member is a LoRA fine-tune of a public 4B or 7B Qwen model trained on the official split alone, with no distillation and no external data. We note that two of the three voting members emit no rationale, so our Track B think field is attached after prediction from a separate reasoning-trained model whose predicted call for that row is identical: it is a post-hoc explanation, not the computation that produced the arguments. We report the experiments that failed alongside those that worked.
%U https://aclanthology.org/2026.arabicnlp-sharedtasks.208/
%P 1608-1613
Markdown (Informal)
[Yalla at AISA-ArabicFC: Argument-Focused Fine-Tuning and Per-Argument Ensembling for Arabic Function Calling](https://aclanthology.org/2026.arabicnlp-sharedtasks.208/) (Joshi, ArabicNLP 2026)
ACL