@inproceedings{aditya-ekbal-2026-adi,
title = "{ADI}-{IITP}: Reward-Guided Preference Optimization for Low-Resource {E}nglish-{A}ssamese Machine Translation",
author = "Aditya, Aditya and
Ekbal, Asif",
editor = "Haddow, Barry and
Kocmi, Tom and
Koehn, Philipp and
Monz, Christof",
booktitle = "Proceedings of the Eleventh Conference on Machine Translation",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.wmt-main.146/",
pages = "2078--2085",
ISBN = "979-8-89176-474-3",
abstract = "This paper describes our submission to the WMT 2026 shared task on Low-Resource Indic Language Translation for the English-Assamese pair. Our systems build on the compact, distilled 200M-parameter IndicTrans2 model, adapted with parameter-efficient Low-Rank Adaptation (LoRA). Beyond standard supervised fine-tuning (SFT), we built a preference-optimization pipeline that generates candidate translations, scores them with a composite reward combining a GEMBA-style LLM-as-judge score, a reference-free CometKiwi quality-estimation signal, and an xCOMET score, and then trains with Direct Preference Optimization (DPO) on the resulting preference pairs. We used this pipeline to guide our system choices: for Assamese-to-English the SFT-only model scored best on our development set, so we submitted it as primary and the SFT-then-DPO model as a contrastive system; for English-to-Assamese we went with SFT-then-DPO as primary. On the official WMT 2026 blind test set, the Assamese-to-English primary system scored 24.08 BLEU (25.11 for the contrastive system), and the English-to-Assamese primary system scored 15.57 BLEU. We also report METEOR, TER, chrF++, BERTScore, and COMET."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="aditya-ekbal-2026-adi">
<titleInfo>
<title>ADI-IITP: Reward-Guided Preference Optimization for Low-Resource English-Assamese Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aditya</namePart>
<namePart type="family">Aditya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Asif</namePart>
<namePart type="family">Ekbal</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>This paper describes our submission to the WMT 2026 shared task on Low-Resource Indic Language Translation for the English-Assamese pair. Our systems build on the compact, distilled 200M-parameter IndicTrans2 model, adapted with parameter-efficient Low-Rank Adaptation (LoRA). Beyond standard supervised fine-tuning (SFT), we built a preference-optimization pipeline that generates candidate translations, scores them with a composite reward combining a GEMBA-style LLM-as-judge score, a reference-free CometKiwi quality-estimation signal, and an xCOMET score, and then trains with Direct Preference Optimization (DPO) on the resulting preference pairs. We used this pipeline to guide our system choices: for Assamese-to-English the SFT-only model scored best on our development set, so we submitted it as primary and the SFT-then-DPO model as a contrastive system; for English-to-Assamese we went with SFT-then-DPO as primary. On the official WMT 2026 blind test set, the Assamese-to-English primary system scored 24.08 BLEU (25.11 for the contrastive system), and the English-to-Assamese primary system scored 15.57 BLEU. We also report METEOR, TER, chrF++, BERTScore, and COMET.</abstract>
<identifier type="citekey">aditya-ekbal-2026-adi</identifier>
<location>
<url>https://aclanthology.org/2026.wmt-main.146/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2078</start>
<end>2085</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ADI-IITP: Reward-Guided Preference Optimization for Low-Resource English-Assamese Machine Translation
%A Aditya, Aditya
%A Ekbal, Asif
%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 aditya-ekbal-2026-adi
%X This paper describes our submission to the WMT 2026 shared task on Low-Resource Indic Language Translation for the English-Assamese pair. Our systems build on the compact, distilled 200M-parameter IndicTrans2 model, adapted with parameter-efficient Low-Rank Adaptation (LoRA). Beyond standard supervised fine-tuning (SFT), we built a preference-optimization pipeline that generates candidate translations, scores them with a composite reward combining a GEMBA-style LLM-as-judge score, a reference-free CometKiwi quality-estimation signal, and an xCOMET score, and then trains with Direct Preference Optimization (DPO) on the resulting preference pairs. We used this pipeline to guide our system choices: for Assamese-to-English the SFT-only model scored best on our development set, so we submitted it as primary and the SFT-then-DPO model as a contrastive system; for English-to-Assamese we went with SFT-then-DPO as primary. On the official WMT 2026 blind test set, the Assamese-to-English primary system scored 24.08 BLEU (25.11 for the contrastive system), and the English-to-Assamese primary system scored 15.57 BLEU. We also report METEOR, TER, chrF++, BERTScore, and COMET.
%U https://aclanthology.org/2026.wmt-main.146/
%P 2078-2085
Markdown (Informal)
[ADI-IITP: Reward-Guided Preference Optimization for Low-Resource English-Assamese Machine Translation](https://aclanthology.org/2026.wmt-main.146/) (Aditya & Ekbal, WMT 2026)
ACL