<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mhaskar-etal-2026-tarl">
    <titleInfo>
        <title>TARL: A Terminology-Aware Agentic Repair Loop for Language-Agnostic Translation</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shivam</namePart>
        <namePart type="family">Mhaskar</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Vrunda</namePart>
        <namePart type="family">Sukhadia</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Anurag</namePart>
        <namePart type="family">Deshmukh</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Manan</namePart>
        <namePart type="family">Sharma</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Pawan</namePart>
        <namePart type="family">Rajpoot</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 Eleventh Conference on Machine Translation</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Barry</namePart>
            <namePart type="family">Haddow</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tom</namePart>
            <namePart type="family">Kocmi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Philipp</namePart>
            <namePart type="family">Koehn</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Christof</namePart>
            <namePart type="family">Monz</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">979-8-89176-474-3</identifier>
    </relatedItem>
    <abstract>We describe TARL (Terminology-Aware Agentic Repair Loop), our submission to the WMT26 Terminology Translation Task, which targets document-level, terminology-constrained translation for morphologically rich and lower-resource directions (Spanish→Basque, English→Polish) and Traditional Chinese→English. Our system performs no fine-tuning: it is a sequential pipeline of four role-specialized LLM agents (translate, verify, review, and post-edit) in which a required target term, once inserted, is treated as locked so that later agents improve the translation around the terminology rather than altering it. Relevant glossary terms are retrieved per sentence by an exact, fuzzy n-gram matcher (with substring matching for CJK source), and for the sample-only Track 2 we first induce a glossary by extracting bilingual term pairs from the provided parallel data. The agents’ instructions are optimized automatically with GEPA (Genetic-Pareto), an automated prompt optimization framework that uses natural language reflection and multi-objective Pareto evolutionary search to tune LLM prompts, rather than being hand-written. On the official WMT26 Track 1 evaluation, TARL attains the highest lemmatized terminology success rate of any submitted system (95.7%, averaged over English→Polish and Spanish→Basque), while remaining competitive on translation quality (COMET-22: 88.3, XCOMET-XXL: 83.6). On Track 2, TARL achieves 80.4% term success and 85.8 COMET-22 across three directions. A terminology-mode analysis confirms that the gains come from genuine use of the supplied dictionary rather than from instruction-following alone.</abstract>
    <identifier type="citekey">mhaskar-etal-2026-tarl</identifier>
    <location>
        <url>https://aclanthology.org/2026.wmt-main.119/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>1840</start>
            <end>1848</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T TARL: A Terminology-Aware Agentic Repair Loop for Language-Agnostic Translation
%A Mhaskar, Shivam
%A Sukhadia, Vrunda
%A Deshmukh, Anurag
%A Sharma, Manan
%A Rajpoot, Pawan
%Y Haddow, Barry
%Y Kocmi, Tom
%Y Koehn, Philipp
%Y Monz, Christof
%S Proceedings of the Eleventh Conference on Machine Translation
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-474-3
%F mhaskar-etal-2026-tarl
%X We describe TARL (Terminology-Aware Agentic Repair Loop), our submission to the WMT26 Terminology Translation Task, which targets document-level, terminology-constrained translation for morphologically rich and lower-resource directions (Spanish→Basque, English→Polish) and Traditional Chinese→English. Our system performs no fine-tuning: it is a sequential pipeline of four role-specialized LLM agents (translate, verify, review, and post-edit) in which a required target term, once inserted, is treated as locked so that later agents improve the translation around the terminology rather than altering it. Relevant glossary terms are retrieved per sentence by an exact, fuzzy n-gram matcher (with substring matching for CJK source), and for the sample-only Track 2 we first induce a glossary by extracting bilingual term pairs from the provided parallel data. The agents’ instructions are optimized automatically with GEPA (Genetic-Pareto), an automated prompt optimization framework that uses natural language reflection and multi-objective Pareto evolutionary search to tune LLM prompts, rather than being hand-written. On the official WMT26 Track 1 evaluation, TARL attains the highest lemmatized terminology success rate of any submitted system (95.7%, averaged over English→Polish and Spanish→Basque), while remaining competitive on translation quality (COMET-22: 88.3, XCOMET-XXL: 83.6). On Track 2, TARL achieves 80.4% term success and 85.8 COMET-22 across three directions. A terminology-mode analysis confirms that the gains come from genuine use of the supplied dictionary rather than from instruction-following alone.
%U https://aclanthology.org/2026.wmt-main.119/
%P 1840-1848

Markdown (Informal)

[TARL: A Terminology-Aware Agentic Repair Loop for Language-Agnostic Translation](https://aclanthology.org/2026.wmt-main.119/) (Mhaskar et al., WMT 2026)

• TARL: A Terminology-Aware Agentic Repair Loop for Language-Agnostic Translation (Mhaskar et al., WMT 2026)

ACL

• Shivam Mhaskar, Vrunda Sukhadia, Anurag Deshmukh, Manan Sharma, and Pawan Rajpoot. 2026. TARL: A Terminology-Aware Agentic Repair Loop for Language-Agnostic Translation. In Proceedings of the Eleventh Conference on Machine Translation, pages 1840–1848, Budapest, Hungary. Association for Computational Linguistics.