This work studies reference-free post-training for multilingual machine translation with open large language models and finds that on-policy distillation reaches, but does not surpass, the quality frontier achieved by RL with checkpoint interpolation.
Abstract
We study reference-free post-training for multilingual machine translation with open large language models. Starting from the supervised-finetuned MiLMMT-46-v0.1 models, we apply Group Relative Policy Optimization (GRPO) with a reward that averages two reference-free quality estimation models and is gated by language identification. We then linearly interpolate the supervised fine-tuning (SFT) and reinforcement learning (RL) model checkpoints to obtain MiLMMT-46-v1.0. Across 46 languages, the resulting models consistently improve translation quality over their SFT counterparts, outperform strong recent open baselines, including Seed-X, HY-MT2, and TranslateGemma, and achieve leading reference-free scores against evaluated proprietary systems such as Google Translate, Gemini 3 Pro, and GPT-5. We further investigate on-policy distillation (OPD) and find that it reaches, but does not surpass, the quality frontier achieved by RL with checkpoint interpolation. We release the models and code to facilitate future research.
Large language models continue to face challenges in translating low-resource languages with scarce parallel data. This study investigates how to fine-tune them effectively using target-side monolingual data. Existing approaches—dominated by back-translation and recent LLM-based rewriting—remain limited by noisy synthetic sources, unguided simplification, and the absence of a principled mechanism for integrating monolingual sentences into the training objective. To address this, we developed a semi-supervised framework that integrates marginal distribution estimation and curriculum-guided rewriting to exploit monolingual data for low-resource translation. Experiments in four low-resource directions demonstrated substantial gains, averaging +8 spBLEU and +10 COMET over strong baselines, while three additional mid-resource directions showed stable improvements and consistent trends. Reference-free metrics further validated robust gains in fluency and adequacy. The findings establish a scalable paradigm for low-resource translation, revealing that the principled integration of marginal likelihood estimation and generative rewriting enables large language models to achieve superior performance under extreme data scarcity.
Wenjie Yu, Zhiqiang Yu, Zuo Jiang et al.· ACM Transactions on Asian an...· 0 citations
This work proposes augmenting existing benchmarks to increase translation difficulty by combining adversarial optimization with a differentiable translation difficulty estimator, and uses gradients from a combined difficulty and fluency objective to iteratively replace tokens in Adversarial Translation Optimization (ATO).
William Kalikman, Šimon Sukup, Michal Tesnar et al.· 0 citations
The results show that multilingual transfer is the dominant factor in extremely low-resource Bantu translation while eliminating the need for heuristic proxy selection, and all systems fail to preserve tonal diacritics, highlighting an open challenge.
Samiratu Ntohsi, Neza David Tuyishimire, Anesu Kafesu et al.· 0 citations
Machine translation (MT) technologies are currently undergoing a paradigm shift, transitioning
from specialized Neural Machine Translation (NMT) frameworks to the broader capabilities of
Large Language Models (LLMs). This paper examines the current standing of the Croatian language
within this technological evolution.
While bilingual NMT models often exhibit high precision, multilingual NMT leverage transfer
learning to enhance performance for low–resource language pairs, but with lower performance for
high–resource ones. Conversely, LLMs—whether general–purpose or fine–tuned for translation—
offer superior multilingual proficiency and context awareness. Unlike NMT, LLMs can process extended discourse, such as full paragraphs or documents, leading to significant improvements in
coreference resolution and gender agreement. Despite the substantial computational requirements
of LLMs, recent optimization techniques allow for smaller, more efficient versions that maintain
high output quality.
This study evaluates the performance of various NMT and LLM architectures specifically for
Croatian from/to English and Spanish using several automatic quality evaluation metrics. The findings demonstrate that open–source models can achieve, and occasionally surpass, the quality of
Google Translate, a widely used commercial NMT system. Furthermore, while our evaluation focuses on this specific language triad, the multilingual nature of the analysed systems suggests that
open–source models provide high–quality translation capabilities for Croatian across dozens, if not
hundreds, of language pairs.
Antoni Oliver, Sergi Álvarez–Vidal· Suvremena Lingvistika· 1 citation
An open end-to-end recipe for training retrieval models and study how English supervision transfers to multilingual retrieval through translate-train are presented, suggesting that token-level matching turns translate-train from a target-language expansion strategy into a multilingual generalization recipe.
Raphaël Sourty, Antoine Chaffin, Paulo Roberto Milanez Oliveira Junior et al.· 1 citation
Large language models (LLMs) are increasingly used as general-purpose translation systems, but their behavior is usually evaluated under a single prompt shape: translate one source sentence into one target language. In practice, users may ask for one target language, for several related languages at once, or for translations conditioned on examples. This paper studies prompt scope and demonstration selection as experimental variables for local LLM machine translation. We evaluate English-to-Romance and English-to-Germanic translation on the full FLORES devtest split for nine official European Union languages. We compare three local instruction-tuned LLMs, llama3.2:3b, mistral:latest, and qwen2.5:14b, against dedicated MT baselines from OPUS-MT and NLLB-200. We test zero-shot prompting and k=5 few-shot prompting with random, lexical-similarity, and embedding-similarity demonstration selection. We also compare single-target prompts with JSON-formatted family-scope prompts that request all languages in a family at once. Results show that dedicated MT systems remain strongest overall, especially for Germanic languages. Few-shot prompting helps mistral:latest and qwen2.5:14b, but hurts llama3.2:3b; embedding retrieval is best on average for the stronger LLMs, but its advantage over random and lexical examples is modest. Family-scope prompting is feasible for stronger local LLMs but exposes structured-output failures in smaller models. These findings motivate evaluating LLM translation not only by language pair and metric, but also by prompt scope, retrieval strategy, and multi-target compliance.