For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, vulnerable to reward-hacking, and provide unactionable assessments. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking objective progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.

User-generated content (UGC) is characterised by frequent use of non-standard language, from spelling errors to expressive choices such as slang, character repetitions, and emojis. This makes evaluating UGC translation challenging: what counts as a "good" translation depends on the desired standardness level of the output. To explore this, we examine the human translation guidelines of four UGC datasets, and derive a taxonomy of twelve non-standard phenomena and five translation actions (NORMALISE, COPY, TRANSFER, OMIT, CENSOR). Our analysis reveals notable differences in how UGC is treated, resulting in a spectrum of standardness in reference translations. We show that translation scores of large language models are highly sensitive to prompts with explicit UGC translation instructions, and that they improve when they align with the dataset guidelines. We argue that fair evaluation requires both models and metrics to be aware of translation guidelines. Finally, we call for clear guidelines during dataset creation and for the development of controllable, guideline-aware evaluation frameworks for UGC translation.

The recent shift from dedicated NMT systems to general-purpose LLMs has reshaped machine translation, with LLMs reported to produce more fluent, less literal output than their predecessors. We test whether this shift extends to the deliteralization hypothesis, the long-standing claim from translation studies that translations become progressively less literal as they are drafted and revised. Using the WMT24++ dataset, we compare the literality of human translations and post-editions to that of two NMT systems and six LLMs across 54 language pairs and three tasks: direct translation, iterative self-revision, and post-editing of human drafts. Literality is measured via a validated Synthetic Literality Index built from six heuristics. We find that (i) human translations remain significantly less literal than those of all tested MT systems, though recent LLMs narrow the gap; (ii) when prompted to iteratively revise their own output, LLMs deliteralize monotonically, providing the first evidence that the hypothesis applies natively to LLM generation; and (iii) as post-editors, LLMs invert the revision triggers of human post-editors, tolerating literal drafts and targeting idiomatic human formulations for revision.

Mechanistic Interpretability (MI) seeks to explain how neural networks implement their capabilities, but the scale of Large Language Models (LLMs) has limited prior MI work in Machine Translation (MT) to word-level analyses. We study sentence-level MT from a mechanistic perspective by analyzing attention heads to understand how LLMs internally encode and distribute translation functions. We decompose MT into two subtasks: producing text in the target language (i.e. target language identification) and preserving the input sentence’s meaning (i.e. sentence equivalence). Across three families of open-source models and 20 translation directions, we find that distinct, sparse sets of attention heads specialize in each subtask. Based on this insight, we construct subtask-specific steering vectors and show that modifying just 1% of the relevant heads enables instruction-free MT performance comparable to instruction-based prompting, while ablating these heads selectively disrupts their corresponding translation functions.

Recent advances in Automatic Text Recognition (ATR) have improved access to historical archives, yet a methodological divide persists between palaeographic transcriptions and normalized digital editions. While ATR models trained on more palaeographically-oriented datasets such as CATMuS have shown greater generalizability, their raw outputs remain poorly compatible with most readers and downstream NLP tools, thus creating a usability gap. On the other hand, ATR models trained to produce normalized outputs have been shown to struggle to adapt to new domains and tend to over-normalize and hallucinate. We introduce the task of Pre-Editorial Normalization (PEN), which consists in normalizing graphemic ATR output according to editorial conventions, which has the advantage of keeping an intermediate step with palaeographic fidelity while providing a normalized version for practical usability. We present a new dataset derived from the CoMMA corpus and aligned with digitized Old French and Latin editions using passim. We also produce a manually corrected gold-standard evaluation set. We benchmark this resource using ByT5-based sequence-to-sequence models on normalization and pre-annotation tasks. Our contributions include the formal definition of PEN, a 4.66M-sample silver training corpus, a 1.8k-sample gold evaluation set, and a normalization model achieving a 6.7

Current human evaluation of machine translation typically assesses single outputs in isolation, a paradigm that suffers from high annotator noise and cost. We introduce Contrastive Error Span Annotation (cESA), a protocol that presents multiple translations of the source input (text, video, audio, image). In cESA, the annotator sees multiple translations of the same document, marks major and minor error spans, and then assigns a score from 0

We release Gaperon, a fully open suite of French-English-coding language models designed to advance transparency and reproducibility in large-scale model training. The Gaperon family includes 1.5B, 8B, and 24B parameter models trained on 2-4 trillion tokens, released with all elements of the training pipeline: French and English datasets filtered with a neural quality classifier, an efficient data curation and training framework, and hundreds of intermediate checkpoints. Through this work, we study how data filtering and contamination interact to shape both benchmark and generative performance. We find that filtering for linguistic quality enhances text fluency and coherence but yields subpar benchmark results, and that late deliberate contamination – continuing training on data mixes that include test sets – recovers competitive scores while only reasonably harming generation quality. We discuss how usual neural filtering can unintentionally amplify benchmark leakage. To support further research, we also introduce harmless data poisoning during pretraining, providing a realistic testbed for safety studies. By openly releasing all models, datasets, code, and checkpoints, Gaperon establishes a reproducible foundation for exploring the trade-offs between data curation, evaluation, safety, and openness in multilingual language model development.

This paper investigates two complementary paradigms for predicting machine translation (MT) quality: source-side difficulty prediction and candidate-side quality estimation (QE). The rapid adoption of Large Language Models (LLMs) into MT workflows is reshaping the research landscape, yet its impact on established quality prediction paradigms remains underexplored. We study this issue through a series of "hindsight" experiments on a unique, multi-candidate dataset resulting from a genuine MT post-editing (MTPE) project. The dataset consists of over 6,000 English source segments with nine translation hypotheses from a diverse set of traditional neural MT systems and advanced LLMs, all evaluated against a single, final human post-edited reference. Using Kendall's rank correlation, we assess the predictive power of source-side difficulty metrics, candidate-side QE models and position heuristics against two gold-standard scores: TER (as a proxy for post-editing effort) and COMET (as a proxy for human judgment). Our findings highlight that the architectural shift towards LLMs alters the reliability of established quality prediction methods while simultaneously mitigating previous challenges in document-level translation.

This study explores an LLM's ability to learn new languages using explanations found in a grammar book, a process we term "explicit learning." To rigorously assess this ability, we design controlled translation experiments between English and constructed languages generated, through specific cryptographic means, from Latin or French. Contrary to previous studies, our results demonstrate that LLMs do possess a measurable capacity for explicit learning. This ability, however, diminishes as the complexity of the linguistic phenomena to be learned increases. Supervised fine-tuning on ad hoc chains of thought significantly enhances LLM performance but struggles to generalize to typologically novel or more complex linguistic features. These findings point to the need for more diverse training sets and alternative fine-tuning strategies to further improve explicit learning by LLMs, benefiting low-resource languages typically described in grammar books but lacking extensive corpora.

The ability of generative large language models (LLMs) to perform in-context learning has given rise to a large body of research into how best to prompt models for various natural language processing tasks. Machine Translation (MT) has been shown to benefit from in-context examples, in particular when they are semantically similar to the sentence to translate. In this paper, we propose a new LLM-based translation paradigm, compositional translation, to replace naive few-shot MT with similarity-based demonstrations. An LLM is used to decompose a sentence into simpler phrases, and then to translate each phrase with the help of retrieved demonstrations. Finally, the LLM is prompted to translate the initial sentence with the help of the self-generated phrase-translation pairs. Our intuition is that this approach should improve translation because these shorter phrases should be intrinsically easier to translate and easier to match with relevant examples. This is especially beneficial in low-resource scenarios, and more generally whenever the selection pool is small or out of domain. We show that compositional translation boosts LLM translation performance on a wide range of popular MT benchmarks, including FLORES200, NTREX 128 and TICO-19. Code and outputs will be made freely available.

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.

Current multimodal machine translation (MMT) systems rely on fully supervised data (i.e models are trained on sentences with their translations and accompanying images). However, this type of data is costly to collect, limiting the extension of MMT to other language pairs for which such data does not exist. In this work, we propose a method to bypass the need for fully supervised data to train MMT systems, using multimodal English data only. Our method, called ZeroMMT, consists in adapting a strong text-only machine translation (MT) model by training it on a mixture of two objectives: visually conditioned masked language modelling and the Kullback-Leibler divergence between the original and new MMT outputs. We evaluate on standard MMT benchmarks and the recently released CoMMuTE, a contrastive benchmark aiming to evaluate how well models use images to disambiguate English sentences. We obtain disambiguation performance close to state-of-the-art MMT models trained additionally on fully supervised examples. To prove that our method generalizes to languages with no fully supervised training data available, we extend the CoMMuTE evaluation dataset to three new languages: Arabic, Russian and Chinese. We further show that we can control the trade-off between disambiguation capabilities and translation fidelity at inference time using classifier-free guidance and without any additional data. Our code, data and trained models are publicly accessible.

This paper introduces AFRIDOC-MT, a document-level multi-parallel translation dataset covering English and five African languages: Amharic, Hausa, Swahili, Yorùbá, and Zulu. The dataset comprises 334 health and 271 information technology news documents, all human-translated from English to these languages. We conduct document-level translation benchmark experiments by evaluating the ability of neural machine translation (NMT) models and large language models (LLMs) to translate between English and these languages, at both the sentence and pseudo-document levels, the outputs being realigned to form complete documents for evaluation. Our results indicate that NLLB-200 achieves the best average performance among the standard NMT models, while GPT-4o outperforms general-purpose LLMs. Fine-tuning selected models leads to substantial performance gains, but models trained on sentences struggle to generalize effectively to longer documents. Furthermore, our analysis reveals that some LLMs exhibit issues such as under-generation, over-generation, repetition of words and phrases, and off-target translations, specifically for translation into African languages.

The ability of generative large language models (LLMs) to perform in-context learning has given rise to a large body of research into how best to prompt models for various natural language processing tasks. In this paper, we focus on machine translation (MT), a task that has been shown to benefit from in-context translation examples. However no systematic studies have been published on how best to select examples, and mixed results have been reported on the usefulness of similarity-based selection over random selection. We provide a study covering multiple LLMs and multiple in-context example retrieval strategies, comparing multilingual sentence embeddings. We cover several language directions, representing different levels of language resourcedness (English into French, German, Swahili and Wolof). Contrarily to previously published results, we find that sentence embedding similarity can improve MT, especially for low-resource language directions, and discuss the balance between selection pool diversity and quality. We also highlight potential problems with the evaluation of LLM-based MT and suggest a more appropriate evaluation protocol, adapting the COMET metric to the evaluation of LLMs. Code and outputs are freely available at https://github.com/ArmelRandy/ICL-MT.

Multimodal Large Language Models (mLLMs) are trained on a large amount of text-image data. While most mLLMs are trained on caption-like data only, Alayrac et al. [2022] showed that additionally training them on interleaved sequences of text and images can lead to the emergence of in-context learning capabilities. However, the dataset they used, M3W, is not public and is only in English. There have been attempts to reproduce their results but the released datasets are English-only. In contrast, current multilingual and multimodal datasets are either composed of caption-like only or medium-scale or fully private data. This limits mLLM research for the 7,000 other languages spoken in the world. We therefore introduce mOSCAR, to the best of our knowledge the first large-scale multilingual and multimodal document corpus crawled from the web. It covers 163 languages, 315M documents, 214B tokens and 1.2B images. We carefully conduct a set of filtering and evaluation steps to make sure mOSCAR is sufficiently safe, diverse and of good quality. We additionally train two types of multilingual model to prove the benefits of mOSCAR: (1) a model trained on a subset of mOSCAR and captioning data and (2) a model train on captioning data only. The model additionally trained on mOSCAR shows a strong boost in few-shot learning performance across various multilingual image-text tasks and benchmarks, confirming previous findings for English-only mLLMs.