Markus Freitag

Freda Shi, Mirac Suzgun, Markus Freitag et al. · deepmind

We evaluate the reasoning abilities of large language models in multilingual settings. We introduce the Multilingual Grade School Math (MGSM) benchmark, by manually translating 250 grade-school math problems from the GSM8K dataset (Cobbe et al., 2021) into ten typologically diverse languages. We find that the ability to solve MGSM problems via chain-of-thought prompting emerges with increasing model scale, and that models have strikingly strong multilingual reasoning abilities, even in underrepresented languages such as Bengali and Swahili. Finally, we show that the multilingual reasoning abilities of language models extend to other tasks such as commonsense reasoning and word-in-context semantic judgment. The MGSM benchmark is publicly available at https://github.com/google-research/url-nlp.

Jingwei Ni, Zhijing Jin, Markus Freitag et al. · deepmind, eth-zurich

Human-translated text displays distinct features from naturally written text in the same language. This phenomena, known as translationese, has been argued to confound the machine translation (MT) evaluation. Yet, we find that existing work on translationese neglects some important factors and the conclusions are mostly correlational but not causal. In this work, we collect CausalMT, a dataset where the MT training data are also labeled with the human translation directions. We inspect two critical factors, the train-test direction match (whether the human translation directions in the training and test sets are aligned), and data-model direction match (whether the model learns in the same direction as the human translation direction in the dataset). We show that these two factors have a large causal effect on the MT performance, in addition to the test-model direction mismatch highlighted by existing work on the impact of translationese. In light of our findings, we provide a set of suggestions for MT training and evaluation. Our code and data are at https://github.com/EdisonNi-hku/CausalMT

Patrick Fernandes, Daniel Deutsch, Mara Finkelstein et al. · deepmind

Automatic evaluation of machine translation (MT) is a critical tool driving the rapid iterative development of MT systems. While considerable progress has been made on estimating a single scalar quality score, current metrics lack the informativeness of more detailed schemes that annotate individual errors, such as Multidimensional Quality Metrics (MQM). In this paper, we help fill this gap by proposing AutoMQM, a prompting technique which leverages the reasoning and in-context learning capabilities of large language models (LLMs) and asks them to identify and categorize errors in translations. We start by evaluating recent LLMs, such as PaLM and PaLM-2, through simple score prediction prompting, and we study the impact of labeled data through in-context learning and finetuning. We then evaluate AutoMQM with PaLM-2 models, and we find that it improves performance compared to just prompting for scores (with particularly large gains for larger models) while providing interpretability through error spans that align with human annotations.

David Vilar, Markus Freitag, Colin Cherry et al. · deepmind, mit

Large language models (LLMs) that have been trained on multilingual but not parallel text exhibit a remarkable ability to translate between languages. We probe this ability in an in-depth study of the pathways language model (PaLM), which has demonstrated the strongest machine translation (MT) performance among similarly-trained LLMs to date. We investigate various strategies for choosing translation examples for few-shot prompting, concluding that example quality is the most important factor. Using optimized prompts, we revisit previous assessments of PaLM's MT capabilities with more recent test sets, modern MT metrics, and human evaluation, and find that its performance, while impressive, still lags that of state-of-the-art supervised systems. We conclude by providing an analysis of PaLM's MT output which reveals some interesting properties and prospects for future work.

Wenda Xu, Daniel Deutsch, Mara Finkelstein et al. · cmu

Recent large language models (LLM) are leveraging human feedback to improve their generation quality. However, human feedback is costly to obtain, especially during inference. In this work, we propose LLMRefine, an inference time optimization method to refine LLM's output. The core idea is to use a learned fine-grained feedback model to pinpoint defects and guide LLM to refine them iteratively. Using original LLM as a proposal of edits, LLMRefine searches for defect-less text via simulated annealing, trading off the exploration and exploitation. We conduct experiments on three text generation tasks, including machine translation, long-form question answering (QA), and topical summarization. LLMRefine consistently outperforms all baseline approaches, achieving improvements up to 1.7 MetricX points on translation tasks, 8.1 ROUGE-L on ASQA, 2.2 ROUGE-L on topical summarization.

Patrick Fernandes, Behrooz Ghorbani, Xavier Garcia et al. · deepmind

In this work, we provide a large-scale empirical study of the scaling properties of multilingual neural machine translation models. We examine how increases in the model size affect the model performance and investigate the role of the training mixture composition on the scaling behavior. We find that changing the weightings of the individual language pairs in the training mixture only affect the multiplicative factor of the scaling law. In particular, we observe that multilingual models trained using different mixing rates all exhibit the same scaling exponent. Through a novel joint scaling law formulation, we compute the effective number of parameters allocated to each language pair and examine the role of language similarity in the scaling behavior of our models. We find little evidence that language similarity has any impact. In contrast, the direction of the multilinguality plays a significant role, with models translating from multiple languages into English having a larger number of effective parameters per task than their reversed counterparts. Finally, we leverage our observations to predict the performance of multilingual models trained with any language weighting at any scale, significantly reducing efforts required for language balancing in large multilingual models. Our findings apply to both in-domain and out-of-domain test sets and to multiple evaluation metrics, such as ChrF and BLEURT.

Tom Kocmi, Eleftherios Avramidis, Rachel Bawden et al. · eth-zurich, microsoft-research

This is the preliminary ranking of WMT24 General MT systems based on automatic metrics. The official ranking will be a human evaluation, which is superior to the automatic ranking and supersedes it. The purpose of this report is not to interpret any findings but only provide preliminary results to the participants of the General MT task that may be useful during the writing of the system submission.

Mara Finkelstein, Subhajit Naskar, Mehdi Mirzazadeh et al. · deepmind

Recent research in decoding methods for Natural Language Generation (NLG) tasks has shown that MAP decoding is not optimal, because model probabilities do not always align with human preferences. Stronger decoding methods, including Quality Estimation (QE) reranking and Minimum Bayes' Risk (MBR) decoding, have since been proposed to mitigate the model-perplexity-vs-quality mismatch. While these decoding methods achieve state-of-the-art performance, they are prohibitively expensive to compute. In this work, we propose MBR finetuning and QE finetuning which distill the quality gains from these decoding methods at training time, while using an efficient decoding algorithm at inference time. Using the canonical NLG task of Neural Machine Translation (NMT), we show that even with self-training, these finetuning methods significantly outperform the base model. Moreover, when using an external LLM as a teacher model, these finetuning methods outperform finetuning on human-generated references. These findings suggest new ways to leverage monolingual data to achieve improvements in model quality that are on par with, or even exceed, improvements from human-curated data, while maintaining maximum efficiency during decoding.

Belén Saldías, George Foster, Markus Freitag et al. · deepmind

Improvements in text generation technologies such as machine translation have necessitated more costly and time-consuming human evaluation procedures to ensure an accurate signal. We investigate a simple way to reduce cost by reducing the number of text segments that must be annotated in order to accurately predict a score for a complete test set. Using a sampling approach, we demonstrate that information from document membership and automatic metrics can help improve estimates compared to a pure random sampling baseline. We achieve gains of up to 20% in average absolute error by leveraging stratified sampling and control variates. Our techniques can improve estimates made from a fixed annotation budget, are easy to implement, and can be applied to any problem with structure similar to the one we study.

Eleftheria Briakou, Jiaming Luo, Colin Cherry et al.

In this paper we present a step-by-step approach to long-form text translation, drawing on established processes in translation studies. Instead of viewing machine translation as a single, monolithic task, we propose a framework that engages language models in a multi-turn interaction, encompassing pre-translation research, drafting, refining, and proofreading, resulting in progressively improved translations. Extensive automatic evaluations using Gemini 1.5 Pro across ten language pairs show that translating step-by-step yields large translation quality improvements over conventional zero-shot prompting approaches and earlier human-like baseline strategies, resulting in state-of-the-art results on WMT2024.

Jan-Thorsten Peter, David Vilar, Daniel Deutsch et al.

Quality Estimation (QE), the evaluation of machine translation output without the need of explicit references, has seen big improvements in the last years with the use of neural metrics. In this paper we analyze the viability of using QE metrics for filtering out bad quality sentence pairs in the training data of neural machine translation systems~(NMT). While most corpus filtering methods are focused on detecting noisy examples in collections of texts, usually huge amounts of web crawled data, QE models are trained to discriminate more fine-grained quality differences. We show that by selecting the highest quality sentence pairs in the training data, we can improve translation quality while reducing the training size by half. We also provide a detailed analysis of the filtering results, which highlights the differences between both approaches.

Daniel Deutsch, Juraj Juraska, Mara Finkelstein et al.

As research on machine translation moves to translating text beyond the sentence level, it remains unclear how effective automatic evaluation metrics are at scoring longer translations. In this work, we first propose a method for creating paragraph-level data for training and meta-evaluating metrics from existing sentence-level data. Then, we use these new datasets to benchmark existing sentence-level metrics as well as train learned metrics at the paragraph level. Interestingly, our experimental results demonstrate that using sentence-level metrics to score entire paragraphs is equally as effective as using a metric designed to work at the paragraph level. We speculate this result can be attributed to properties of the task of reference-based evaluation as well as limitations of our datasets with respect to capturing all types of phenomena that occur in paragraph-level translations.

Mara Finkelstein, David Vilar, Markus Freitag

Recent research in neural machine translation (NMT) has shown that training on high-quality machine-generated data can outperform training on human-generated data. This work accompanies the first-ever release of a LLM-generated, MBR-decoded and QE-reranked dataset with both sentence-level and multi-sentence examples. We perform extensive experiments to demonstrate the quality of our dataset in terms of its downstream impact on NMT model performance. We find that training from scratch on our (machine-generated) dataset outperforms training on the (web-crawled) WMT'23 training dataset (which is 300 times larger), and also outperforms training on the top-quality subset of the WMT'23 training dataset. We also find that performing self-distillation by finetuning the LLM which generated this dataset outperforms the LLM's strong few-shot baseline. These findings corroborate the quality of our dataset, and demonstrate the value of high-quality machine-generated data in improving performance of NMT models.

Christian Tomani, David Vilar, Markus Freitag et al.

Maximum-a-posteriori (MAP) decoding is the most widely used decoding strategy for neural machine translation (NMT) models. The underlying assumption is that model probability correlates well with human judgment, with better translations getting assigned a higher score by the model. However, research has shown that this assumption does not always hold, and generation quality can be improved by decoding to optimize a utility function backed by a metric or quality-estimation signal, as is done by Minimum Bayes Risk (MBR) or quality-aware decoding. The main disadvantage of these approaches is that they require an additional model to calculate the utility function during decoding, significantly increasing the computational cost. In this paper, we propose to make the NMT models themselves quality-aware by training them to estimate the quality of their own output. Using this approach for MBR decoding we can drastically reduce the size of the candidate list, resulting in a speed-up of two-orders of magnitude. When applying our method to MAP decoding we obtain quality gains similar or even superior to quality reranking approaches, but with the efficiency of single pass decoding.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Jan-Thorsten Peter, David Vilar, Tobias Domhan et al.

Most current large language models (LLMs) support a wide variety of languages in addition to English, including high-resource languages (e.g. German, Chinese, French), as well as low-resource ones (e.g. Swahili, Telugu). In addition they have also shown impressive capabilities in different domains, like coding, science and math. In this short paper, taking math as an example domain, we study the performance of different LLMs across languages. Experimental results show that there exists a non-negligible and consistent gap in the performance of the models across languages. Interestingly, and somewhat against expectations, the gap exists for both high- and low-resource languages. We hope that these results influence further research into cross-lingual capability generalization for next generation LLMs. If it weren't for the fact that they are false! By analyzing one of the standard multilingual math benchmarks (MGSM), we determine that several translation errors are present in the data. Furthermore, the lack of standardized answer extraction from LLM outputs further influences the final results. We propose a method for automatic quality assurance to address the first issue at scale, and give recommendations to address the second one. Combining these two approaches we show that the aforementioned language gap mostly disappears, leading to completely different conclusions from our research. We additionally release the corrected dataset to the community.

Daniel Deutsch, Eleftheria Briakou, Isaac Caswell et al.

As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, including on tasks like machine translation (MT). In this work, we extend the WMT24 dataset to cover 55 languages by collecting new human-written references and post-edits for 46 new languages and dialects in addition to post-edits of the references in 8 out of 9 languages in the original WMT24 dataset. The dataset covers four domains: literary, news, social, and speech. We benchmark a variety of MT providers and LLMs on the collected dataset using automatic metrics and find that LLMs are the best-performing MT systems in all 55 languages. These results should be confirmed using a human-based evaluation, which we leave for future work.

Geza Kovacs, Daniel Deutsch, Markus Freitag

While Minimum Bayes Risk (MBR) decoding using metrics such as COMET or MetricX has outperformed traditional decoding methods such as greedy or beam search, it introduces a challenge we refer to as metric bias. As MBR decoding aims to produce translations that score highly according to a specific utility metric, this very process makes it impossible to use the same metric for both decoding and evaluation, as improvements might simply be due to reward hacking rather than reflecting real quality improvements. In this work we find that compared to human ratings, neural metrics not only overestimate the quality of MBR decoding when the same metric is used as the utility metric, but they also overestimate the quality of MBR/QE decoding with other neural utility metrics as well. We also show that the metric bias issue can be mitigated by using an ensemble of utility metrics during MBR decoding: human evaluations show that MBR decoding using an ensemble of utility metrics outperforms a single utility metric.

Parker Riley, Daniel Deutsch, George Foster et al.

Reliable human evaluation is critical to the development of successful natural language generation models, but achieving it is notoriously difficult. Stability is a crucial requirement when ranking systems by quality: consistent ranking of systems across repeated evaluations is not just desirable, but essential. Without it, there is no reliable foundation for hill-climbing or product launch decisions. In this paper, we use machine translation and its state-of-the-art human evaluation framework, MQM, as a case study to understand how to set up reliable human evaluations that yield stable conclusions. We investigate the optimal configurations for item allocation to raters, number of ratings per item, and score normalization. Our study on two language pairs provides concrete recommendations for designing replicable human evaluation studies. We also collect and release the largest publicly available dataset of multi-segment translations rated by multiple professional translators, consisting of nearly 140,000 segment annotations across two language pairs.

Gergely Flamich, David Vilar, Jan-Thorsten Peter et al.

The goal of translation, be it by human or by machine, is, given some text in a source language, to produce text in a target language that simultaneously 1) preserves the meaning of the source text and 2) achieves natural expression in the target language. However, researchers in the machine translation community usually assess translations using a single score intended to capture semantic accuracy and the naturalness of the output simultaneously. In this paper, we build on recent advances in information theory to mathematically prove and empirically demonstrate that such single-score summaries do not and cannot give the complete picture of a system's true performance. Concretely, we prove that a tradeoff exists between accuracy and naturalness and demonstrate it by evaluating the submissions to the WMT24 shared task. Our findings help explain well-known empirical phenomena, such as the observation that optimizing translation systems for a specific accuracy metric (like BLEU) initially improves the system's naturalness, while ``overfitting'' the system to the metric can significantly degrade its naturalness. Thus, we advocate for a change in how translations are evaluated: rather than comparing systems using a single number, they should be compared on an accuracy-naturalness plane.

Muhammed Yusuf Kocyigit, Eleftheria Briakou, Daniel Deutsch et al. · mit

Data contamination -- the accidental consumption of evaluation examples within the pre-training data -- can undermine the validity of evaluation benchmarks. In this paper, we present a rigorous analysis of the effects of contamination on language models at 1B and 8B scales on the machine translation task. Starting from a carefully decontaminated train-test split, we systematically introduce contamination at various stages, scales, and data formats to isolate its effect and measure its impact on performance metrics. Our experiments reveal that contamination with both source and target substantially inflates BLEU scores, and this inflation is 2.5 times larger (up to 30 BLEU points) for 8B compared to 1B models. In contrast, source-only and target-only contamination generally produce smaller, less consistent over-estimations. Finally, we study how the temporal distribution and frequency of contaminated samples influence performance over-estimation across languages with varying degrees of data resources.

Zhongtao Liu, Parker Riley, Daniel Deutsch et al.

Collecting high-quality translations is crucial for the development and evaluation of machine translation systems. However, traditional human-only approaches are costly and slow. This study presents a comprehensive investigation of 11 approaches for acquiring translation data, including human-only, machineonly, and hybrid approaches. Our findings demonstrate that human-machine collaboration can match or even exceed the quality of human-only translations, while being more cost-efficient. Error analysis reveals the complementary strengths between human and machine contributions, highlighting the effectiveness of collaborative methods. Cost analysis further demonstrates the economic benefits of human-machine collaboration methods, with some approaches achieving top-tier quality at around 60% of the cost of traditional methods. We release a publicly available dataset containing nearly 18,000 segments of varying translation quality with corresponding human ratings to facilitate future research.

Yixiao Song, Parker Riley, Daniel Deutsch et al.

Human evaluation is crucial for assessing rapidly evolving language models but is influenced by annotator proficiency and task design. This study explores the integration of comparative judgment into human annotation for machine translation (MT) and evaluates three annotation setups-point-wise Multidimensional Quality Metrics (MQM), side-by-side (SxS) MQM, and its simplified version SxS relative ranking (RR). In MQM, annotators mark error spans with categories and severity levels. SxS MQM extends MQM to pairwise error annotation for two translations of the same input, while SxS RR focuses on selecting the better output without labeling errors. Key findings are: (1) the SxS settings achieve higher inter-annotator agreement than MQM; (2) SxS MQM enhances inter-translation error marking consistency compared to MQM by, on average, 38.5% for explicitly compared MT systems and 19.5% for others; (3) all annotation settings return stable system rankings, with SxS RR offering a more efficient alternative to (SxS) MQM; (4) the SxS settings highlight subtle errors overlooked in MQM without altering absolute system evaluations. To spur further research, we will release the triply annotated datasets comprising 377 ZhEn and 104 EnDe annotation examples.

Mara Finkelstein, Isaac Caswell, Tobias Domhan et al.

We present TranslateGemma, a suite of open machine translation models based on the Gemma 3 foundation models. To enhance the inherent multilingual capabilities of Gemma 3 for the translation task, we employ a two-stage fine-tuning process. First, supervised fine-tuning is performed using a rich mixture of high-quality large-scale synthetic parallel data generated via state-of-the-art models and human-translated parallel data. This is followed by a reinforcement learning phase, where we optimize translation quality using an ensemble of reward models, including MetricX-QE and AutoMQM, targeting translation quality. We demonstrate the effectiveness of TranslateGemma with human evaluation on the WMT25 test set across 10 language pairs and with automatic evaluation on the WMT24++ benchmark across 55 language pairs. Automatic metrics show consistent and substantial gains over the baseline Gemma 3 models across all sizes. Notably, smaller TranslateGemma models often achieve performance comparable to larger baseline models, offering improved efficiency. We also show that TranslateGemma models retain strong multimodal capabilities, with enhanced performance on the Vistra image translation benchmark. The release of the open TranslateGemma models aims to provide the research community with powerful and adaptable tools for machine translation.

Mara Finkelstein, Dan Deutsch, Parker Riley et al.

As LLMs continue to become more powerful and versatile, human evaluation has quickly become intractable at scale and reliance on automatic metrics has become the norm. Recently, it has been shown that LLMs are themselves state-of-the-art evaluators for many tasks. These Autoraters are typically designed so that they generalize to new systems and test sets. In practice, however, evaluation is performed on a small set of fixed, canonical test sets, which are carefully curated to measure certain capabilities of interest and are not changed frequently. In this work, we design a method which specializes a prompted Autorater to a given test set, by leveraging historical ratings on the test set to construct in-context learning (ICL) examples. We evaluate our Specialist method on the task of fine-grained machine translation evaluation, and show that it dramatically outperforms the state-of-the-art XCOMET metric by 54% and 119% on the WMT'23 and WMT'24 test sets, respectively. We perform extensive analyses to understand the representations learned by our Specialist metrics, and how variability in rater behavior affects their performance. We also verify the generalizability and robustness of our Specialist method for designing automatic metrics across different numbers of ICL examples, LLM backbones, systems to evaluate, and evaluation tasks.

Lily H. Zhang, Hamid Dadkhahi, Mara Finkelstein et al. · mit

Despite growing interest in incorporating feedback to improve language models, most efforts focus only on sequence-level annotations. In this work, we explore the potential of utilizing fine-grained span-level annotations from offline datasets to improve model quality. We develop a simple finetuning algorithm, called Training with Annotations (TWA), to directly train machine translation models on such annotated data. TWA utilizes targeted span-level error information while also flexibly learning what to penalize within a span. Moreover, TWA considers the overall trajectory of a sequence when deciding which non-error spans to utilize as positive signals. Experiments on English-German and Chinese-English machine translation show that TWA outperforms baselines such as Supervised FineTuning on sequences filtered for quality and Direct Preference Optimization on pairs constructed from the same data.

Juraj Juraska, Tobias Domhan, Mara Finkelstein et al.

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.

Vilém Zouhar, Wenda Xu, Parker Riley et al. · eth-zurich

Machine translation benchmarks sourced from the real world are quickly obsoleted, due to most examples being easy for state-of-the-art translation models. This limits the benchmark's ability to distinguish which model is better or to reveal models' weaknesses. Current methods for creating difficult test cases, such as subsampling or from-scratch synthesis, either fall short of identifying difficult examples or suffer from a lack of diversity and naturalness. Inspired by the iterative process of human experts probing for model failures, we propose MT-breaker, a method where a large language model iteratively refines a source text to increase its translation difficulty. The LLM iteratively queries a target machine translation model to guide its generation of difficult examples. Our approach generates examples that are more challenging for the target MT model while preserving the diversity of natural texts. While the examples are tailored to a particular machine translation model during the generation, the difficulty also transfers to other models and languages.

Parker Riley, Daniel Deutsch, Mara Finkelstein et al.

Human evaluation of machine translation is in an arms race with translation model quality: as our models get better, our evaluation methods need to be improved to ensure that quality gains are not lost in evaluation noise. To this end, we experiment with a two-stage version of the current state-of-the-art translation evaluation paradigm (MQM), which we call MQM re-annotation. In this setup, an MQM annotator reviews and edits a set of pre-existing MQM annotations, that may have come from themselves, another human annotator, or an automatic MQM annotation system. We demonstrate that rater behavior in re-annotation aligns with our goals, and that re-annotation results in higher-quality annotations, mostly due to finding errors that were missed during the first pass.

Yilin Zhang, Wenda Xu, Zhongtao Liu et al.

Quality Estimation (QE) metrics are vital in machine translation for reference-free evaluation and as a reward signal in tasks like reinforcement learning. However, the prevalence and impact of length bias in QE have been underexplored. Through a systematic study of top-performing regression-based and LLM-as-a-Judge QE metrics across 10 diverse language pairs, we reveal two critical length biases: First, QE metrics consistently over-predict errors with increasing translation length, even for high-quality, error-free texts. Second, they exhibit a preference for shorter translations when multiple candidates are available for the same source text. These inherent length biases risk unfairly penalizing longer, correct translations and can lead to sub-optimal decision-making in applications such as QE reranking and QE guided reinforcement learning. To mitigate this, we propose two strategies: (a) applying length normalization during model training, and (b) incorporating reference texts during evaluation. Both approaches were found to effectively reduce the identified length bias.

Wenda Xu, Vilém Zouhar, Parker Riley et al. · eth-zurich

NLP models require test data that are sufficiently challenging. The difficulty of an example is linked to the topic it originates from (''seed topic''). The relationship between the topic and the difficulty of its instances is stochastic in nature: an example about a difficult topic can happen to be easy, and vice versa. At the scale of the Internet, there are tens of thousands of potential topics, and finding the most difficult one by drawing and evaluating a large number of examples across all topics is computationally infeasible. We formalize this task and treat it as a multi-armed bandit problem. In this framework, each topic is an ''arm,'' and pulling an arm (at a cost) involves drawing a single example, evaluating it, and measuring its difficulty. The goal is to efficiently identify the most difficult topics within a fixed computational budget. We illustrate the bandit problem setup of finding difficult examples for the task of machine translation. We find that various bandit strategies vastly outperform baseline methods like brute-force searching the most challenging topics.

Wenda Xu, Sweta Agrawal, Vilém Zouhar et al. · eth-zurich

As large language models (LLMs) begin to saturate existing benchmarks, automated benchmark creation using LLMs (LLM as a benchmark) has emerged as a scalable alternative to slow and costly human curation. While these generated test sets have to potential to cheaply rank models, we demonstrate a critical flaw. LLM generated benchmarks systematically favor the model that created the benchmark, they exhibit self bias on low resource languages to English translation tasks. We show three key findings on automatic benchmarking of LLMs for translation: First, this bias originates from two sources: the generated test data (LLM as a testset) and the evaluation method (LLM as an evaluator), with their combination amplifying the effect. Second, self bias in LLM as a benchmark is heavily influenced by the model's generation capabilities in the source language. For instance, we observe more pronounced bias in into English translation, where the model's generation system is developed, than in out of English translation tasks. Third, we observe that low diversity in source text is one attribution to self bias. Our results suggest that improving the diversity of these generated source texts can mitigate some of the observed self bias.

Behzad Shayegh, Jan-Thorsten Peter, David Vilar et al.

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.

Tom Kocmi, Eleftherios Avramidis, Rachel Bawden et al. · eth-zurich, microsoft-research

We present the preliminary rankings of machine translation (MT) systems submitted to the WMT25 General Machine Translation Shared Task, as determined by automatic evaluation metrics. Because these rankings are derived from automatic evaluation, they may exhibit a bias toward systems that employ re-ranking techniques, such as Quality Estimation or Minimum Bayes Risk decoding. The official WMT25 ranking will be based on human evaluation, which is more reliable and will supersede these results. The official WMT25 ranking will be based on human evaluation, which is more reliable and will supersede these results. The purpose of releasing these findings now is to assist task participants with their system description papers; not to provide final findings.

Firas Trabelsi, David Vilar, Mara Finkelstein et al.

Minimum Bayes Risk (MBR) decoding is a powerful decoding strategy widely used for text generation tasks, but its quadratic computational complexity limits its practical application. This paper presents a novel approach for approximating MBR decoding using matrix completion techniques, focusing on the task of machine translation. We formulate MBR decoding as a matrix completion problem, where the utility metric scores between candidate hypotheses and pseudo-reference translations form a low-rank matrix. First, we empirically show that the scores matrices indeed have a low-rank structure. Then, we exploit this by only computing a random subset of the scores and efficiently recover the missing entries in the matrix by applying the Alternating Least Squares (ALS) algorithm, thereby enabling a fast approximation of the MBR decoding process. Our experimental results on machine translation tasks demonstrate that the proposed method requires 1/16 utility metric computations compared to vanilla MBR decoding while achieving equal translation quality measured by COMET22 on the WMT22 dataset (en<>de and en<>ru). We also benchmark our method against other approximation methods and we show gains in quality when comparing to them.

Daniel Deutsch, George Foster, Markus Freitag

Kendall's tau is frequently used to meta-evaluate how well machine translation (MT) evaluation metrics score individual translations. Its focus on pairwise score comparisons is intuitive but raises the question of how ties should be handled, a gray area that has motivated different variants in the literature. We demonstrate that, in settings like modern MT meta-evaluation, existing variants have weaknesses arising from their handling of ties, and in some situations can even be gamed. We propose instead to meta-evaluate metrics with a version of pairwise accuracy that gives metrics credit for correctly predicting ties, in combination with a tie calibration procedure that automatically introduces ties into metric scores, enabling fair comparison between metrics that do and do not predict ties. We argue and provide experimental evidence that these modifications lead to fairer ranking-based assessments of metric performance.

Wenda Xu, Danqing Wang, Liangming Pan et al.

Automatically evaluating the quality of language generation is critical. Although recent learned metrics show high correlation with human judgement, these metrics can not explain their verdict or associate the scores with defects in generated text. To address this limitation, we present InstructScore, an explainable evaluation metric for text generation. By harnessing both explicit human instruction and the implicit knowledge of GPT-4, we fine-tune a text evaluation metric based on LLaMA, producing both a score for generated text and a human readable diagnostic report. We evaluate InstructScore on a variety of generation tasks, including translation, captioning, data-to-text and commonsense generation. Experiments show that our 7B model surpasses all other unsupervised metrics, including those based on 175B GPT-3 and GPT-4. Surprisingly, our InstructScore, even without direct supervision from human-rated data, achieves performance levels on par with state-of-the-art metrics like COMET22, which were fine-tuned on human ratings.

Rohan Anil, Andrew M. Dai, Orhan Firat et al.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Markus Freitag, Behrooz Ghorbani, Patrick Fernandes

Recent advances in machine translation (MT) have shown that Minimum Bayes Risk (MBR) decoding can be a powerful alternative to beam search decoding, especially when combined with neural-based utility functions. However, the performance of MBR decoding depends heavily on how and how many candidates are sampled from the model. In this paper, we explore how different sampling approaches for generating candidate lists for MBR decoding affect performance. We evaluate popular sampling approaches, such as ancestral, nucleus, and top-k sampling. Based on our insights into their limitations, we experiment with the recently proposed epsilon-sampling approach, which prunes away all tokens with a probability smaller than epsilon, ensuring that each token in a sample receives a fair probability mass. Through extensive human evaluations, we demonstrate that MBR decoding based on epsilon-sampling significantly outperforms not only beam search decoding, but also MBR decoding with all other tested sampling methods across four language pairs.

Markus Freitag, David Grangier, Qijun Tan et al.

In Neural Machine Translation, it is typically assumed that the sentence with the highest estimated probability should also be the translation with the highest quality as measured by humans. In this work, we question this assumption and show that model estimates and translation quality only vaguely correlate. We apply Minimum Bayes Risk (MBR) decoding on unbiased samples to optimize diverse automated metrics of translation quality as an alternative inference strategy to beam search. Instead of targeting the hypotheses with the highest model probability, MBR decoding extracts the hypotheses with the highest estimated quality. Our experiments show that the combination of a neural translation model with a neural reference-based metric, BLEURT, results in significant improvement in human evaluations. This improvement is obtained with translations different from classical beam-search output: these translations have much lower model likelihood and are less favored by surface metrics like BLEU.

Behrooz Ghorbani, Orhan Firat, Markus Freitag et al.

We present an empirical study of scaling properties of encoder-decoder Transformer models used in neural machine translation (NMT). We show that cross-entropy loss as a function of model size follows a certain scaling law. Specifically (i) We propose a formula which describes the scaling behavior of cross-entropy loss as a bivariate function of encoder and decoder size, and show that it gives accurate predictions under a variety of scaling approaches and languages; we show that the total number of parameters alone is not sufficient for such purposes. (ii) We observe different power law exponents when scaling the decoder vs scaling the encoder, and provide recommendations for optimal allocation of encoder/decoder capacity based on this observation. (iii) We also report that the scaling behavior of the model is acutely influenced by composition bias of the train/test sets, which we define as any deviation from naturally generated text (either via machine generated or human translated text). We observe that natural text on the target side enjoys scaling, which manifests as successful reduction of the cross-entropy loss. (iv) Finally, we investigate the relationship between the cross-entropy loss and the quality of the generated translations. We find two different behaviors, depending on the nature of the test data. For test sets which were originally translated from target language to source language, both loss and BLEU score improve as model size increases. In contrast, for test sets originally translated from source language to target language, the loss improves, but the BLEU score stops improving after a certain threshold. We release generated text from all models used in this study.

Scott Roy, Cliff Brunk, Kyu-Young Kim et al.

One of the challenges in a task oriented natural language application like the Google Assistant, Siri, or Alexa is to localize the output to many languages. This paper explores doing this by applying machine translation to the English output. Using machine translation is very scalable, as it can work with any English output and can handle dynamic text, but otherwise the problem is a poor fit. The required quality bar is close to perfection, the range of sentences is extremely narrow, and the sentences are often very different than the ones in the machine translation training data. This combination of requirements is novel in the field of domain adaptation for machine translation. We are able to reach the required quality bar by building on existing ideas and adding new ones: finetuning on in-domain translations, adding sentences from the Web, adding semantic annotations, and using automatic error detection. The paper shares our approach and results, together with a distillation model to serve the translation models at scale.

Kelly Marchisio, Markus Freitag, David Grangier

Modern unsupervised machine translation (MT) systems reach reasonable translation quality under clean and controlled data conditions. As the performance gap between supervised and unsupervised MT narrows, it is interesting to ask whether the different training methods result in systematically different output beyond what is visible via quality metrics like adequacy or BLEU. We compare translations from supervised and unsupervised MT systems of similar quality, finding that unsupervised output is more fluent and more structurally different in comparison to human translation than is supervised MT. We then demonstrate a way to combine the benefits of both methods into a single system which results in improved adequacy and fluency as rated by human evaluators. Our results open the door to interesting discussions about how supervised and unsupervised MT might be different yet mutually-beneficial.

Markus Freitag, George Foster, David Grangier et al.

Human evaluation of modern high-quality machine translation systems is a difficult problem, and there is increasing evidence that inadequate evaluation procedures can lead to erroneous conclusions. While there has been considerable research on human evaluation, the field still lacks a commonly-accepted standard procedure. As a step toward this goal, we propose an evaluation methodology grounded in explicit error analysis, based on the Multidimensional Quality Metrics (MQM) framework. We carry out the largest MQM research study to date, scoring the outputs of top systems from the WMT 2020 shared task in two language pairs using annotations provided by professional translators with access to full document context. We analyze the resulting data extensively, finding among other results a substantially different ranking of evaluated systems from the one established by the WMT crowd workers, exhibiting a clear preference for human over machine output. Surprisingly, we also find that automatic metrics based on pre-trained embeddings can outperform human crowd workers. We make our corpus publicly available for further research.

Sweta Agrawal, George Foster, Markus Freitag et al.

Reference-free evaluation has the potential to make machine translation evaluation substantially more scalable, allowing us to pivot easily to new languages or domains. It has been recently shown that the probabilities given by a large, multilingual model can achieve state of the art results when used as a reference-free metric. We experiment with various modifications to this model and demonstrate that by scaling it up we can match the performance of BLEU. We analyze various potential weaknesses of the approach and find that it is surprisingly robust and likely to offer reasonable performance across a broad spectrum of domains and different system qualities.

Markus Freitag, George Foster, David Grangier et al.

Automatic evaluation comparing candidate translations to human-generated paraphrases of reference translations has recently been proposed by Freitag et al. When used in place of original references, the paraphrased versions produce metric scores that correlate better with human judgment. This effect holds for a variety of different automatic metrics, and tends to favor natural formulations over more literal (translationese) ones. In this paper we compare the results of performing end-to-end system development using standard and paraphrased references. With state-of-the-art English-German NMT components, we show that tuning to paraphrased references produces a system that is significantly better according to human judgment, but 5 BLEU points worse when tested on standard references. Our work confirms the finding that paraphrased references yield metric scores that correlate better with human judgment, and demonstrates for the first time that using these scores for system development can lead to significant improvements.

Markus Freitag, Orhan Firat

Multilingual Neural Machine Translation (MNMT) models are commonly trained on a joint set of bilingual corpora which is acutely English-centric (i.e. English either as the source or target language). While direct data between two languages that are non-English is explicitly available at times, its use is not common. In this paper, we first take a step back and look at the commonly used bilingual corpora (WMT), and resurface the existence and importance of implicit structure that existed in it: multi-way alignment across examples (the same sentence in more than two languages). We set out to study the use of multi-way aligned examples to enrich the original English-centric parallel corpora. We reintroduce this direct parallel data from multi-way aligned corpora between all source and target languages. By doing so, the English-centric graph expands into a complete graph, every language pair being connected. We call MNMT with such connectivity pattern complete Multilingual Neural Machine Translation (cMNMT) and demonstrate its utility and efficacy with a series of experiments and analysis. In combination with a novel training data sampling strategy that is conditioned on the target language only, cMNMT yields competitive translation quality for all language pairs. We further study the size effect of multi-way aligned data, its transfer learning capabilities and how it eases adding a new language in MNMT. Finally, we stress test cMNMT at scale and demonstrate that we can train a cMNMT model with up to 111*112=12,432 language pairs that provides competitive translation quality for all language pairs.

Thibault Sellam, Amy Pu, Hyung Won Chung et al.

The quality of machine translation systems has dramatically improved over the last decade, and as a result, evaluation has become an increasingly challenging problem. This paper describes our contribution to the WMT 2020 Metrics Shared Task, the main benchmark for automatic evaluation of translation. We make several submissions based on BLEURT, a previously published metric based on transfer learning. We extend the metric beyond English and evaluate it on 14 language pairs for which fine-tuning data is available, as well as 4 "zero-shot" language pairs, for which we have no labelled examples. Additionally, we focus on English to German and demonstrate how to combine BLEURT's predictions with those of YiSi and use alternative reference translations to enhance the performance. Empirical results show that the models achieve competitive results on the WMT Metrics 2019 Shared Task, indicating their promise for the 2020 edition.

Zorik Gekhman, Roee Aharoni, Genady Beryozkin et al.

We propose a simple and effective method for machine translation evaluation which does not require reference translations. Our approach is based on (1) grounding the entity mentions found in each source sentence and candidate translation against a large-scale multilingual knowledge base, and (2) measuring the recall of the grounded entities found in the candidate vs. those found in the source. Our approach achieves the highest correlation with human judgements on 9 out of the 18 language pairs from the WMT19 benchmark for evaluation without references, which is the largest number of wins for a single evaluation method on this task. On 4 language pairs, we also achieve higher correlation with human judgements than BLEU. To foster further research, we release a dataset containing 1.8 million grounded entity mentions across 18 language pairs from the WMT19 metrics track data.

Markus Freitag, David Grangier, Isaac Caswell

The quality of automatic metrics for machine translation has been increasingly called into question, especially for high-quality systems. This paper demonstrates that, while choice of metric is important, the nature of the references is also critical. We study different methods to collect references and compare their value in automated evaluation by reporting correlation with human evaluation for a variety of systems and metrics. Motivated by the finding that typical references exhibit poor diversity, concentrating around translationese language, we develop a paraphrasing task for linguists to perform on existing reference translations, which counteracts this bias. Our method yields higher correlation with human judgment not only for the submissions of WMT 2019 English to German, but also for Back-translation and APE augmented MT output, which have been shown to have low correlation with automatic metrics using standard references. We demonstrate that our methodology improves correlation with all modern evaluation metrics we look at, including embedding-based methods. To complete this picture, we reveal that multi-reference BLEU does not improve the correlation for high quality output, and present an alternative multi-reference formulation that is more effective.