Accurately evaluating machine-translated text remains a long-standing challenge, particularly for long documents. Recent work has shown that large language models (LLMs) can serve as reliable and interpretable sentence-level translation evaluators via MQM error span annotations. With modern LLMs supporting larger context windows, a natural question arises: can we feed entire document translations into an LLM for quality assessment? Ideally, evaluation should be invariant to text length, producing consistent error spans regardless of input granularity. However, our analysis shows that text length significantly impacts evaluation: longer texts lead to fewer error spans and reduced system ranking accuracy. To address this limitation, we evaluate several strategies, including granularity-aligned prompting, Focus Sentence Prompting (FSP), and a fine-tuning approach to better align LLMs with the evaluation task. The latter two methods largely mitigate this length bias, making LLMs more reliable for long-form translation evaluation.

We investigate the tradeoff between adequacy and fluency in machine translation. We show the severity of this tradeoff at the evaluation level and analyze where popular metrics fall within it. Essentially, current metrics generally lean toward adequacy, meaning that their scores correlate more strongly with the adequacy of translations than with fluency. More importantly, we find that this tradeoff also persists at the meta-evaluation level, and that the standard WMT meta-evaluation favors adequacy-oriented metrics over fluency-oriented ones. We show that this bias is partially attributed to the composition of the systems included in the meta-evaluation datasets. To control this bias, we propose a method that synthesizes translation systems in meta-evaluation. Our findings highlight the importance of understanding this tradeoff in meta-evaluation and its impact on metric rankings.

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.

In this paper, we present our submissions to the unified WMT25 Translation Evaluation Shared Task. For the Quality Score Prediction subtask, we create a new generation of MetricX with improvements in the input format and the training protocol, while for the Error Span Detection subtask we develop a new model, GemSpanEval, trained to predict error spans along with their severities and categories. Both systems are based on the state-of-the-art multilingual open-weights model Gemma 3, fine-tuned on publicly available WMT data. We demonstrate that MetricX-25, adapting Gemma 3 to an encoder-only architecture with a regression head on top, can be trained to effectively predict both MQM and ESA quality scores, and significantly outperforms its predecessor. Our decoder-only GemSpanEval model, on the other hand, we show to be competitive in error span detection with xCOMET, a strong encoder-only sequence-tagging baseline. With error span detection formulated as a generative task, we instruct the model to also output the context for each predicted error span, thus ensuring that error spans are identified unambiguously.

We show that content on the web is often translated into many languages, and the low quality of these multi-way translations indicates they were likely created using Machine Translation (MT). Multi-way parallel, machine generated content not only dominates the translations in lower resource languages; it also constitutes a large fraction of the total web content in those languages. We also find evidence of a selection bias in the type of content which is translated into many languages, consistent with low quality English content being translated en masse into many lower resource languages, via MT. Our work raises serious concerns about training models such as multilingual large language models on both monolingual and bilingual data scraped from the web.

Building upon prior WMT shared tasks in document alignment and sentence filtering, we posed the open-ended shared task of finding the best subset of possible training data from a collection of Estonian-Lithuanian web data. Participants could focus on any portion of the end-to-end data curation pipeline, including alignment and filtering. We evaluated results based on downstream machine translation quality. We release processed Common Crawl data, along with various intermediate states from a strong baseline system, which we believe will enable future research on this topic.

Neural metrics trained on human evaluations of MT tend to correlate well with human judgments, but their behavior is not fully understood. In this paper, we perform a controlled experiment and compare a baseline metric that has not been trained on human evaluations (Prism) to a trained version of the same metric (Prism+FT). Surprisingly, we find that Prism+FT becomes more robust to machine-translated references, which are a notorious problem in MT evaluation. This suggests that the effects of metric training go beyond the intended effect of improving overall correlation with human judgments.

Vocabulary selection, or lexical shortlisting, is a well-known technique to improve latency of Neural Machine Translation models by constraining the set of allowed output words during inference. The chosen set is typically determined by separately trained alignment model parameters, independent of the source-sentence context at inference time. While vocabulary selection appears competitive with respect to automatic quality metrics in prior work, we show that it can fail to select the right set of output words, particularly for semantically non-compositional linguistic phenomena such as idiomatic expressions, leading to reduced translation quality as perceived by humans. Trading off latency for quality by increasing the size of the allowed set is often not an option in real-world scenarios. We propose a model of vocabulary selection, integrated into the neural translation model, that predicts the set of allowed output words from contextualized encoder representations. This restores translation quality of an unconstrained system, as measured by human evaluations on WMT newstest2020 and idiomatic expressions, at an inference latency competitive with alignment-based selection using aggressive thresholds, thereby removing the dependency on separately trained alignment models.

Building neural machine translation systems to perform well on a specific target domain is a well-studied problem. Optimizing system performance for multiple, diverse target domains however remains a challenge. We study this problem in an adaptation setting where the goal is to preserve the existing system quality while incorporating data for domains that were not the focus of the original translation system. We find that we can improve over the performance trade-off offered by Elastic Weight Consolidation with a relatively simple data mixing strategy. At comparable performance on the new domains, catastrophic forgetting is mitigated significantly on strong WMT baselines. Combining both approaches improves the Pareto frontier on this task.

We present Sockeye 2, a modernized and streamlined version of the Sockeye neural machine translation (NMT) toolkit. New features include a simplified code base through the use of MXNet’s Gluon API, a focus on state of the art model architectures, and distributed mixed precision training. These improvements result in faster training and inference, higher automatic metric scores, and a shorter path from research to production.

With recent advances in network architectures for Neural Machine Translation (NMT) recurrent models have effectively been replaced by either convolutional or self-attentional approaches, such as in the Transformer. While the main innovation of the Transformer architecture is its use of self-attentional layers, there are several other aspects, such as attention with multiple heads and the use of many attention layers, that distinguish the model from previous baselines. In this work we take a fine-grained look at the different architectures for NMT. We introduce an Architecture Definition Language (ADL) allowing for a flexible combination of common building blocks. Making use of this language we show in experiments that one can bring recurrent and convolutional models very close to the Transformer performance by borrowing concepts from the Transformer architecture, but not using self-attention. Additionally, we find that self-attention is much more important on the encoder side than on the decoder side, where it can be replaced by a RNN or CNN without a loss in performance in most settings. Surprisingly, even a model without any target side self-attention performs well.

The performance of Neural Machine Translation (NMT) models relies heavily on the availability of sufficient amounts of parallel data, and an efficient and effective way of leveraging the vastly available amounts of monolingual data has yet to be found. We propose to modify the decoder in a neural sequence-to-sequence model to enable multi-task learning for two strongly related tasks: target-side language modeling and translation. The decoder predicts the next target word through two channels, a target-side language model on the lowest layer, and an attentional recurrent model which is conditioned on the source representation. This architecture allows joint training on both large amounts of monolingual and moderate amounts of bilingual data to improve NMT performance. Initial results in the news domain for three language pairs show moderate but consistent improvements over a baseline trained on bilingual data only.