Large Language Models have shown impressive multilingual capabilities, where translation is one among many tasks. Google Translate's submission to the 2025 WMT evaluation tries to research how these models behave when pushing their translation performance to the limit. Starting with the strong Gemma 3 model, we carry out supervised fine tuning on high quality, synthetically generated parallel data. Afterwards we perform an additional Reinforcement Learning step, with reward models based on translation metrics to push the translation capabilities even further. Controlling the combination of reward models, including reference-based and quality estimation metrics, we found that the behaviour of the model could be tailored towards a more literal or more creative translation style. Our two submissions correspond to those two models. We chose the more creative system as our primary submission, targetting a human preference for better sounding, more naturally flowing text, although at the risk of losing on the accuracy of the translation. It is an open question to find the sweet spot between these two dimensions, which certainly will depend on the specific domain to handle and user preferences.
Large Language Models (LLMs) have been demonstrated to achieve state-of-the-art results on machine translation. LLM-based translation systems usually rely on model adaptation and fine-tuning, requiring datasets and compute. The goal of our team's participation in the "General Machine Translation" and "Multilingual" tasks of WMT25 was to evaluate the translation effectiveness of a resource-efficient solution consisting of a smaller off-the-shelf LLM coupled with a self-refine agentic workflow. Our approach requires a high-quality multilingual LLM capable of instruction following. We select Gemma3-12B among several candidates using the pretrained translation metric MetricX-24-XL and a small development dataset. WMT25 automatic evaluations place our solution in the mid tier of all WMT25 systems, and also demonstrate that it can perform competitively for approximately 16% of language pairs.
This paper presents the submission from Dalian University of Technology (DLUT) and Global Tone Communication Technology Co., Ltd. (GTCOM) to the WMT25 General Machine Translation Task. Amidst the paradigm shift from specialized encoder-decoder models to general-purpose Large Language Models (LLMs), this work conducts a systematic comparison of both approaches across five language pairs. For traditional Neural Machine Translation (NMT), we build strong baselines using deep Transformer architectures enhanced with data augmentation. For the LLM paradigm, we explore zero-shot performance and two distinct supervised fine-tuning (SFT) strategies: direct translation and translation refinement. Our key findings reveal a significant discrepancy between lexical and semantic evaluation metrics: while strong NMT systems remain competitive in BLEU scores, fine-tuned LLMs demonstrate marked superiority in semantic fidelity as measured by COMET. Furthermore, we find that fine-tuning LLMs for direct translation is more effective than for refinement, suggesting that teaching the core task directly is preferable to correcting baseline outputs.
This study tackles the WMT 2025 low-resource Indic language translation task (English-Assamese, English-Manipuri) by proposing a cross-iterative back-translation and data augmentation strategy using dual pretrained models.Leveraging IndicTrans2_1B and NLLB_3.3B, the approach alternates fine-tuning and back-translation to iteratively generate high-quality pseudo-parallel corpora. Monolingual data relevance is enhanced via semantic similarity filtering with all-mpnet-basev2, while training data is cleaned and normalized to improve quality. During inference, outputs from both fine-tuned models are combined to further boost translation performance in low-resource scenarios.
This paper describes the TartuNLP submission to the Upper Sorbian (hsb) and Lower Sorbian (dsb) tracks of the WMT25 LLMs with Limited Resources for Slavic Languages shared task, which jointly targets machine translation (MT) and question answering (QA). We develop a single multilingual model based on Qwen2.5-3B-Instruct by continuing pretraining on Sorbian monolingual and parallel data together with general instruction datasets, combining language acquisition and instructionfollowing in a single step. The resulting model delivers substantial improvements over the baseline Qwen2.5-3B-Instruct model and also achieves the highest ranking for both tasks in the hsb and dsb shared task tracks.
We present an English-to-Japanese translation system built upon the EuroLLM-9B (Martins et al., 2025) model. The training process involves two main stages: continue pretraining (CPT) and supervised fine-tuning (SFT). After both stages, we further tuned the model using a development set to optimize performance. For training data, we employed both basic filtering techniques and high-quality filtering strategies to ensure data cleanness. Additionally, we classify both the training data and development data into four different domains and we train and fine-tune with domain specific prompts during system training. Finally, we applied Minimum Bayes Risk (MBR) decoding and paragraph-level reranking for post-processing to enhance translation quality(1).
This paper presents a submission to the WMT25 test suite subtask, focusing on terminology in official documents related to the EEA Agreement. The test suite evaluates the accuracy of MT systems in translating terminology for the English -> Icelandic translation direction when applied to EEA documents. We focus on the use of terminology in four domains of the agreement; science, technology, economics, and society. We find that manual evaluation confirms that our test suite can be helpful in selecting the best MT system for working with these domains. Surprisingly, an online system which does not achieve very high scores on the general translation task, according to preliminary results, is most adept at translating the terminology our test suite evaluates. The test suite and evaluation code are openly available on Github: https://github.com/steinst/WMT25_EEA_test_suite
This paper presents an English to Bhojpuri machine translation (MT) system developed for the WMT25 General MT Shared Task. Given the low-resource nature of Bhojpuri, we adopt a two-stage training pipeline: unsupervised pretraining followed by supervised fine-tuning. During pretraining, we use a 300,000-sentence corpus comprising 70% Bhojpuri monolingual data and 30% English data to establish language grounding. The fine-tuning stage utilizes 29,749 bilingual English to Bhojpuri sentence pairs (including training, validation, and test sets). To adapt the system to instruction-following scenarios, we apply a novel optimization strategy: Contrastive Preference Optimization (CPO). This technique enables the model to capture fine-grained translation preferences and maintain semantic fidelity in instructiontuned settings. We evaluate our system across multiple metrics, demonstrating moderate performance in low-resource MT tasks, particularly in diverse domains such as literary, news, social, and speech.
This year, we focus exclusively on using the uncertainty quantification as a proxy for translation quality. While this has traditionally been regarded as a form of unsupervised quality estimation, such signals have been overlooked in the design of the current metric models-we show their value in the context of LLMs. More specifically, in contrast to conventional unsupervised QE methods, we apply recent calibration technology (Wu et al., 2025b) to adjust translation likelihoods to better align with quality signals, and we use the single resulting model to participate in both the general translation and QE tracks at WMT25. Our offline experiments show some advantages: 1) uncertainty signals extracted from LLMs, like Tower or Gemma-3, provide accurate quality predictions; and 2) calibration technology further improves this QE performance, sometimes even surpassing certain metric models that were trained with human annotations, such as CometKiwi. We therefore argue that uncertainty quantification (confidence), especially from LLMs, can serve as a strong and complementary signal for the metric design, particularly when human-annotated data are lacking. However, we also identify limitations, i.e., its tendency to assign disproportionately higher scores to hypotheses generated by the model itself.
This paper presents UvA-MT's submission to the WMT 2025 shared task on general machine translation, competing in the unconstrained track across all 16 translation directions. Unusually, this year we use only WMT25's blind test set (source sentences only) to generate synthetic data for LLM training, and translations are produced using pure beam search for submission. Overall, our approach can be seen as a special variant of data distillation, motivated by two key considerations: (1) perfect domain alignment, where the training and test domains are distributionally identical; and (2) the strong teacher model, GPT-4o-mini, offers high-quality outputs as both a reliable reference and a fallback in case of mere memorization. Interestingly, the outputs of the resulting model, trained on Gemma3-12B using Best-of-N (BoN) outputs from GPT-4o-mini, outperform both original BoN outputs from GPT-4omini and Gemma3-12B in some high-resource languages across various metrics. We attribute this to a successful model ensemble, where the student model (Gemma3-12B) retains the strengths of the teacher (GPT-4o-mini) while implicitly avoiding its flaws.
Document-level machine translation is a challenging task, as it requires modeling both short-range and long-range dependencies to maintain the coherence and cohesion of the generated translation. However, these dependencies are sparse, and most context-augmented translation systems resort to two equally unsatisfactory options: either to include maximally long contexts, hoping that the useful dependencies are not lost in the noise; or to use limited local contexts, at the risk of missing relevant information. In this work, we study a self-retrieval-augmented machine translation framework (SELF- RAMT), aimed at informing translation decisions with informative local and global contexts dynamically extracted from the source and target texts. We examine the effectiveness of this method using three large language models, considering three criteria for context selection. We carry out experiments on TED talks as well as parallel scientific articles, considering three translation directions. Our results show that integrating distant contexts with SELF-RAMT improves translation quality as measured by reference-based scores and consistency metrics.
This paper describes the Kaze-MT submission to the WMT25 General Machine Translation task (Japanese-Chinese). Our system deliberately adopts a minimalist Test-Time Scaling (TTS) pipeline with three stages-Sampling, Scoring, and Selection-while avoiding any task-specific fine-tuning, in-context exemplars, or bespoke decoding heuristics. In the sampling stage, we use the zero-shot Qwen2.5-72B-Instruct model to generate 512 candidate translations under a fixed temperature schedule designed to encourage lexical and syntactic diversity without sacrificing fluency. In the scoring stage, each candidate is evaluated by multiple reference-free quality estimation (QE) models-KIWI-22, MetricX24 Hybrid-XXL, and Remedy-24-9B. The selection stage aggregates metric-specific rankings and chooses the candidate with the lowest mean rank, which we found more stable than averaging raw scores across heterogeneous ranges. We submit to both constrained and unconstrained tracks with minimal configuration changes. According to official preliminary results, our submissions are competitive on automatic metrics; in human evaluation, Kaze-MT falls within the 8-13 cluster, delivering performance comparable to CommandA-WMT and DeepSeek-V3 and outperforming other large LLM baselines such as Mistral-Medium and other extensively tuned MT systems.
This work describes Laniqo's submission to the constrained track of the WMT25 General MT Task. We participated in 11 translation directions. Our approach combines several techniques: fine-tuning the EuroLLM-9B-Instruct model using Contrastive Preference Optimization on a synthetic dataset, applying Retrieval-Augmented Translation with human-translated data, implementing Quality-Aware Decoding, and performing postprocessing of translations with a rule-based algorithm. We analyze the contribution of each method and report improvements at every stage of our pipeline.
We analyze how English-Russian machine translation (MT) systems submitted to WMT25 perform on linguistically challenging translation tasks, similar to problems used in university professional translator training. We assessed the ten top-performing systems using a fine-grained test suite containing 465 manually devised test items, which cover 55 lexical, grammatical, and discourse phenomena, in 13 categories. By applying pass/fail rules with human adjudication and micro/macro aggregates, we observe three performance tiers. Compared with the official WMT25 ranking, our ranking broadly aligns but reveals notable shifts. Our findings show that in 2025, even topperforming MT systems still struggle with translation problems that require deep understanding and rephrasing, much like human novices do. The best systems exhibit creativity and can be very good at handling such challenges, often producing more natural translations rather than producing word-for-word renditions. However, persistent structural and lexical problems remain: literal word order carryovers, misused verb forms, and rigid phrase translations were common, mirroring errors typically seen in beginner translator assignments.
This work is a submission to the Creole MT Task conducted as a part of the Tenth Conference on Machine Translation (WMT '25) colocated with EMNLP 2025. The focus of the work was on the Seychellois Creole- English language pair by utilizing similar Creoles to make up for the scarcity of data in Creole language systems.
When fine-tuning massively multilingual translation models for low-resource languages, practitioners often include auxiliary languages to improve performance, but factors determining successful auxiliary language selection remain unclear. This paper investigates whether syntactic similarity or lexical overlap is more important for effective multilingual fine-tuning. We use encipherment to create controlled experimental conditions that disentangle these confounded factors, generating novel languages with identical syntax but no lexical overlap, and conversely, languages that preserve lexical overlap. Through extensive NLLB-200 finetuning experiments across Europarl and AmericasNLP datasets, we demonstrate that lexical overlap is the dominant factor. Syntactically identical auxiliary languages provide negligible benefits (< 1.0 ChrF), while languages with significant lexical overlap provide substantial improvements (> 5.0 ChrF), with effectiveness strongly correlated to KL-divergence between token distributions (r = -0.47, p <.001). Our findings provide clear guidance: when selecting auxiliary languages for multilingual finetuning, prioritize lexical overlap over syntactic similarity.
We describe Vicomtech's participation in the WMT 2025 Shared Task on Model Compression. We addressed all three language pairs of the constrained task, namely Czech to German, English to Arabic and Japanese to Chinese, using the Aya Expanse 8B model as our base model. Our approach centers on GeLaCo, an evolutionary method for LLM compression via layer collapse operations, which efficiently explores the compression solution space through populationbased search and a module-wise similarity fitness function that captures attention, feed-forward, and hidden state representations. We systematically evaluated compression at three different ratios (0.25, 0.50, and 0.75) and applied targeted post-training techniques to recover performance through fine-tuning and knowledge distillation over translation instructions. Additionally, we explored quantization techniques to achieve further model size reduction. Our experimental results demonstrate that the combination of evolutionary layer compression, targeted post-training, and quantization can achieve substantial model size reduction while maintaining competitive translation quality across all language pairs.
Intuitively, machine translation (MT) between closely related languages, such as Swedish and Danish, is easier than MT between more distant pairs, such as Finnish and Danish. Yet, the notions of 'closely related' languages and 'easier' translation have so far remained under-specified. Moreover, in the context of neural MT, this assumption was almost exclusively evaluated in scenarios where English was either the source or target language, leaving a broader cross-lingual view unexplored. In this work, we present a controlled study of language similarity and neural MT difficulty for 56 European translation directions. We test a range of language similarity metrics, some of which are reasonable predictors of MT difficulty. On a text-level, we reassess previously introduced indicators of MT difficulty, and find that they are not well-suited to our domain, or neural MT more generally. Ultimately, we hope that this work inspires further cross-lingual investigations of neural MT difficulty.
We present the results of the first edition of the Model Compression shared task, organized as part of the 10th Conference on Machine Translation (WMT25). The task challenged participants to compress Large Language Models (LLMs) toward enabling practical deployment in resource-constrained scenarios, while minimizing loss in translation performance. In this edition, participants could choose to compete in either a constrained track, which required compressing a specific model (Aya Expanse 8B) evaluated on a limited set of language pairs (Czech -> German, Japanese -> Chinese, and English -> Arabic), or an unconstrained track, which placed no restrictions on the model and allowed submissions for any of the 15 language directions covered by the General MT task. We received 12 submissions from 3 teams, all in the constrained track. They proposed different compression solutions and covered various language combinations. Evaluation was conducted separately for each language, measuring translation quality using COMET and MetricX, model size, and inference speed on an Nvidia A100 GPU.
We present the results of the WMT 2025 shared task of the Open Language Data Initiative (OLDI). Participants were invited to contribute to the existing massively multilingual open datasets supported by OLDI (FLORES+, OLDI-Seed, WMT24++), or create new resources in line with OLDI's aims. We accepted eight submissions: seven extensions or revisions of the existing datasets and one submission with a new massively parallel training dataset, SMOL. These contributions advance the coverage and quality of multilingual datasets, especially since for many languages, they are the first publicly-available training or evaluation data for machine translation. All contributions are released under permissive open-source licenses.
