@inproceedings{lee-etal-2026-rl,
title = "{RL} from Teacher-Model Refinement: Gradual Imitation Learning for Machine Translation",
author = "Lee, Dongyub Jude and
Ye, Zhenyi and
He, Pengcheng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.591/",
pages = "12473--12485",
ISBN = "979-8-89176-462-0",
abstract = "Preference-learning methods for machine translation (MT), such as Direct Preference Optimization (DPO), have shown strong gains but typically rely on carefully curated preference triplets and often struggle to generalize beyond their tuning domains. We propose Reinforcement Learning from Teacher-Model Refinement (RLfR), which replaces static triplets with on-policy, actor-conditioned refinements produced by a frozen teacher. At each step, the actor samples candidate translations, the teacher performs a minimal local edit of each draft, and the actor is reinforced to close the gap using a composite reward that combines scaled negative edit distance for lexical and structural fidelity with COMET for semantic adequacy. This formulation yields a stable, model-aware learning signal without requiring curated preference triplets, making it naturally scalable to many language pairs without additional data collection. Experiments on FLORES-200 (en {\ensuremath{<}}-{\ensuremath{>}} de/es/zh/ko/ja) show that RLfR consistently outperforms strong baselines, improving semantic quality and entity preservation across automatic, LLM-judge, and human evaluations."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-etal-2026-rl">
<titleInfo>
<title>RL from Teacher-Model Refinement: Gradual Imitation Learning for Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dongyub</namePart>
<namePart type="given">Jude</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenyi</namePart>
<namePart type="family">Ye</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengcheng</namePart>
<namePart type="family">He</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>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Preference-learning methods for machine translation (MT), such as Direct Preference Optimization (DPO), have shown strong gains but typically rely on carefully curated preference triplets and often struggle to generalize beyond their tuning domains. We propose Reinforcement Learning from Teacher-Model Refinement (RLfR), which replaces static triplets with on-policy, actor-conditioned refinements produced by a frozen teacher. At each step, the actor samples candidate translations, the teacher performs a minimal local edit of each draft, and the actor is reinforced to close the gap using a composite reward that combines scaled negative edit distance for lexical and structural fidelity with COMET for semantic adequacy. This formulation yields a stable, model-aware learning signal without requiring curated preference triplets, making it naturally scalable to many language pairs without additional data collection. Experiments on FLORES-200 (en \ensuremath<-\ensuremath> de/es/zh/ko/ja) show that RLfR consistently outperforms strong baselines, improving semantic quality and entity preservation across automatic, LLM-judge, and human evaluations.</abstract>
<identifier type="citekey">lee-etal-2026-rl</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.591/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12473</start>
<end>12485</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T RL from Teacher-Model Refinement: Gradual Imitation Learning for Machine Translation
%A Lee, Dongyub Jude
%A Ye, Zhenyi
%A He, Pengcheng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F lee-etal-2026-rl
%X Preference-learning methods for machine translation (MT), such as Direct Preference Optimization (DPO), have shown strong gains but typically rely on carefully curated preference triplets and often struggle to generalize beyond their tuning domains. We propose Reinforcement Learning from Teacher-Model Refinement (RLfR), which replaces static triplets with on-policy, actor-conditioned refinements produced by a frozen teacher. At each step, the actor samples candidate translations, the teacher performs a minimal local edit of each draft, and the actor is reinforced to close the gap using a composite reward that combines scaled negative edit distance for lexical and structural fidelity with COMET for semantic adequacy. This formulation yields a stable, model-aware learning signal without requiring curated preference triplets, making it naturally scalable to many language pairs without additional data collection. Experiments on FLORES-200 (en \ensuremath<-\ensuremath> de/es/zh/ko/ja) show that RLfR consistently outperforms strong baselines, improving semantic quality and entity preservation across automatic, LLM-judge, and human evaluations.
%U https://aclanthology.org/2026.findings-emnlp.591/
%P 12473-12485
Markdown (Informal)
[RL from Teacher-Model Refinement: Gradual Imitation Learning for Machine Translation](https://aclanthology.org/2026.findings-emnlp.591/) (Lee et al., Findings 2026)
ACL