Lyrics translation requires both accurate semantic transfer and preservation of musical rhythm, syllabic structure, and poetic style. In animated musicals, the challenge intensifies due to alignment with visual and auditory cues. We introduce Multilingual Audio-Video Lyrics Benchmark for Animated Song Translation (MAVL), the first multilingual, multimodal benchmark for singable lyrics translation. By integrating text, audio, and video, MAVL enables richer and more expressive translations than text-only approaches. Building on this, we propose Syllable-Constrained Audio-Video LLM with Chain-of-Thought SylAVL-CoT, which leverages audio-video cues and enforces syllabic constraints to produce natural-sounding lyrics. Experimental results demonstrate that SylAVL-CoT significantly outperforms text-based models in singability and contextual accuracy, emphasizing the value of multimodal, multilingual approaches for lyrics translation.
Large language models (LLMs) have demonstrated strong performance in general-purpose machine translation, but their effectiveness in complex, domain-sensitive translation tasks remains underexplored. Recent advancements in Large Reasoning Models (LRMs), raise the question of whether structured reasoning can enhance translation quality across diverse domains. In this work, we compare the performance of LRMs with traditional LLMs across 15 representative domains and four translation directions. Our evaluation considers various factors, including task difficulty, input length, and terminology density. We use a combination of automatic metrics and an enhanced MQM-based evaluation hierarchy to assess translation quality. Our findings show that LRMs consistently outperform traditional LLMs in semantically complex domains, especially in long-text and high-difficulty translation scenarios. Moreover, domain-adaptive prompting strategies further improve performance by better leveraging the reasoning capabilities of LRMs. These results highlight the potential of structured reasoning in MDMT tasks and provide valuable insights for optimizing translation systems in domain-sensitive contexts.
Large language models (LLMs) have recently demonstrated remarkable capabilities in machine translation (MT). However, most advanced MT-specific LLMs heavily rely on external supervision signals during training, such as human-annotated reference data or trained reward models (RMs), which are often expensive to obtain and challenging to scale. To overcome this limitation, we propose a Simple Self-Rewarding (SSR) Reinforcement Learning (RL) framework for MT that is reference-free, fully online, and relies solely on self-judging rewards. Training with SSR using 13K monolingual examples and Qwen-2.5-7B as the backbone, our model SSR-Zero-7B outperforms existing MT-specific LLMs, e.g., TowerInstruct-13B and GemmaX-28-9B, as well as larger general LLMs like Qwen2.5-32B-Instruct in English $\leftrightarrow$ Chinese translation tasks from WMT23, WMT24, and Flores200 benchmarks. Furthermore, by augmenting SSR with external supervision from COMET, our strongest model, SSR-X-Zero-7B, achieves state-of-the-art performance in English $\leftrightarrow$ Chinese translation, surpassing all existing open-source models under 72B parameters and even outperforming closed-source models, e.g., GPT-4o and Gemini 1.5 Pro. Our analysis highlights the effectiveness of the self-rewarding mechanism compared to the external LLM-as-a-judge approach in MT and demonstrates its complementary benefits when combined with trained RMs. Our findings provide valuable insight into the potential of self-improving RL methods. We have publicly released our code, data and models.
High-quality machine translation systems based on large language models (LLMs) have simplified the production of personalized translations reflecting specific stylistic constraints. However, these systems still struggle in settings where stylistic requirements are less explicit and might be harder to convey via prompting. We explore various strategies for personalizing LLM-generated translations in low-resource settings, focusing on the challenging literary translation domain. We explore prompting strategies and inference-time interventions for steering model generations towards a personalized style, and propose a contrastive framework exploiting latent concepts extracted from sparse autoencoders to identify salient personalization properties. Our results show that steering achieves strong personalization while preserving translation quality. We further examine the impact of steering on LLM representations, finding model layers with a relevant impact for personalization are impacted similarly by multi-shot prompting and our steering method, suggesting similar mechanism at play.
Recent advances in Automatic Speech Recognition (ASR) have been largely fueled by massive speech corpora. However, extending coverage to diverse languages with limited resources remains a formidable challenge. This paper introduces Speech Back-Translation, a scalable pipeline that improves multilingual ASR models by converting large-scale text corpora into synthetic speech via off-the-shelf text-to-speech (TTS) models. We demonstrate that just tens of hours of real transcribed speech can effectively train TTS models to generate synthetic speech at hundreds of times the original volume while maintaining high quality. To evaluate synthetic speech quality, we develop an intelligibility-based assessment framework and establish clear thresholds for when synthetic data benefits ASR training. Using Speech Back-Translation, we generate more than 500,000 hours of synthetic speech in ten languages and continue pre-training Whisper-large-v3, achieving average transcription error reductions of over 30\%. These results highlight the scalability and effectiveness of Speech Back-Translation for enhancing multilingual ASR systems.
We introduce CASS, the first large-scale dataset and model suite for cross-architecture GPU code transpilation, targeting both source-level (CUDA <--> HIP) and assembly-level (Nvidia SASS <--> AMD RDNA3) translation. The dataset comprises 70k verified code pairs across host and device, addressing a critical gap in low-level GPU code portability. Leveraging this resource, we train the CASS family of domain-specific language models, achieving 95% source translation accuracy and 37.5% assembly translation accuracy, substantially outperforming commercial baselines such as GPT-4o, Claude, and Hipify. Our generated code matches native performance in over 85% of test cases, preserving runtime and memory behavior. To support rigorous evaluation, we introduce CASS-Bench, a curated benchmark spanning 16 GPU domains with ground-truth execution. All data, models, and evaluation tools are released as open source to foster progress in GPU compiler tooling, binary compatibility, and LLM-guided hardware translation.
The rapid use of large language models (LLMs) has raised critical concerns regarding the factual reliability of their outputs, especially in low-resource languages such as Urdu. Existing automated fact-checking solutions overwhelmingly focus on English, leaving a significant gap for the 200+ million Urdu speakers worldwide. In this work, we introduce UrduFactCheck, the first comprehensive, modular fact-checking framework specifically tailored for Urdu. Our system features a dynamic, multi-strategy evidence retrieval pipeline that combines monolingual and translation-based approaches to address the scarcity of high-quality Urdu evidence. We curate and release two new hand-annotated benchmarks: UrduFactBench for claim verification and UrduFactQA for evaluating LLM factuality. Extensive experiments demonstrate that UrduFactCheck, particularly its translation-augmented variants, consistently outperforms baselines and open-source alternatives on multiple metrics. We further benchmark twelve state-of-the-art (SOTA) LLMs on factual question answering in Urdu, highlighting persistent gaps between proprietary and open-source models. UrduFactCheck's code and datasets are open-sourced and publicly available at https://github.com/mbzuai-nlp/UrduFactCheck.
We introduce a data-driven approach for enabling word-level timestamp prediction in the Canary model. Accurate timestamp information is crucial for a variety of downstream tasks such as speech content retrieval and timed subtitles. While traditional hybrid systems and end-to-end (E2E) models may employ external modules for timestamp prediction, our approach eliminates the need for separate alignment mechanisms. By leveraging the NeMo Forced Aligner (NFA) as a teacher model, we generate word-level timestamps and train the Canary model to predict timestamps directly. We introduce a new <|timestamp|> token, enabling the Canary model to predict start and end timestamps for each word. Our method demonstrates precision and recall rates between 80% and 90%, with timestamp prediction errors ranging from 20 to 120 ms across four languages, with minimal WER degradation. Additionally, we extend our system to automatic speech translation (AST) tasks, achieving timestamp prediction errors around 200 milliseconds.
We present MAATS, a Multi Agent Automated Translation System that leverages the Multidimensional Quality Metrics (MQM) framework as a fine-grained signal for error detection and refinement. MAATS employs multiple specialized AI agents, each focused on a distinct MQM category (e.g., Accuracy, Fluency, Style, Terminology), followed by a synthesis agent that integrates the annotations to iteratively refine translations. This design contrasts with conventional single-agent methods that rely on self-correction. Evaluated across diverse language pairs and Large Language Models (LLMs), MAATS outperforms zero-shot and single-agent baselines with statistically significant gains in both automatic metrics and human assessments. It excels particularly in semantic accuracy, locale adaptation, and linguistically distant language pairs. Qualitative analysis highlights its strengths in multi-layered error diagnosis, omission detection across perspectives, and context-aware refinement. By aligning modular agent roles with interpretable MQM dimensions, MAATS narrows the gap between black-box LLMs and human translation workflows, shifting focus from surface fluency to deeper semantic and contextual fidelity.
In-Image Machine Translation (IIMT) aims to translate texts within images from one language to another. Previous research on IIMT was primarily conducted on simplified scenarios such as images of one-line text with black font in white backgrounds, which is far from reality and impractical for applications in the real world. To make IIMT research practically valuable, it is essential to consider a complex scenario where the text backgrounds are derived from real-world images. To facilitate research of complex scenario IIMT, we design an IIMT dataset that includes subtitle text with real-world background. However previous IIMT models perform inadequately in complex scenarios. To address the issue, we propose the DebackX model, which separates the background and text-image from the source image, performs translation on text-image directly, and fuses the translated text-image with the background, to generate the target image. Experimental results show that our model achieves improvements in both translation quality and visual effect.
While Transformer self-attention offers strong parallelism, the Key-Value (KV) cache grows linearly with sequence length and becomes a bottleneck for inference efficiency. Multi-head latent attention was recently developed to compress the KV cache into a low-rank latent space. This paper proposes Multi-head Temporal Latent Attention (MTLA), which further reduces the KV cache size along the temporal dimension, greatly lowering the memory footprint of self-attention inference. MTLA employs a hyper-network to dynamically merge temporally adjacent KV cache vectors. To address the mismatch between the compressed KV cache and processed sequence lengths, a stride-aware causal mask is proposed to ensure efficient parallel training and consistency with inference behaviour. Experiments across tasks, including speech translation, speech recognition, speech understanding and text summarisation, demonstrate that MTLA achieves competitive performance compared to standard Multi-Head Attention (MHA), while greatly improving inference speed and GPU memory usage. For example, on a English-German speech translation task, MTLA achieves a 5.3x speedup and a reduction in GPU memory usage by a factor of 8.3 compared to MHA, while maintaining translation quality.
The rapid advancement of large vision-language models (LVLMs) has significantly propelled applications in document understanding, particularly in optical character recognition (OCR) and multilingual translation. However, current evaluations of LVLMs, like the widely used OCRBench, mainly focus on verifying the correctness of their short-text responses and long-text responses with simple layout, while the evaluation of their ability to understand long texts with complex layout design is highly significant but largely overlooked. In this paper, we propose Menu OCR and Translation Benchmark (MOTBench), a specialized evaluation framework emphasizing the pivotal role of menu translation in cross-cultural communication. MOTBench requires LVLMs to accurately recognize and translate each dish, along with its price and unit items on a menu, providing a comprehensive assessment of their visual understanding and language processing capabilities. Our benchmark is comprised of a collection of Chinese and English menus, characterized by intricate layouts, a variety of fonts, and culturally specific elements across different languages, along with precise human annotations. Experiments show that our automatic evaluation results are highly consistent with professional human evaluation. We evaluate a range of publicly available state-of-the-art LVLMs, and through analyzing their output to identify the strengths and weaknesses in their performance, offering valuable insights to guide future advancements in LVLM development. MOTBench is available at https://github.com/gitwzl/MOTBench.
In recent years, the emergence of large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, has shown impressive capabilities in complex problems, e.g., mathematics and coding. Some pioneering studies attempt to bring the success of LRMs in neural machine translation (MT). They try to build LRMs with deep reasoning MT ability via reinforcement learning (RL). Despite some progress that has been made, these attempts generally focus on several high-resource languages, e.g., English and Chinese, leaving the performance on other languages unclear. Besides, the reward modeling methods in previous work do not fully unleash the potential of reinforcement learning in MT. In this work, we first design a new reward modeling method that compares the translation results of the policy MT model with a strong LRM (i.e., DeepSeek-R1-671B), and quantifies the comparisons to provide rewards. Experimental results demonstrate the superiority of the reward modeling method. Using Qwen2.5-7B-Instruct as the backbone, the trained model achieves the new state-of-the-art performance in literary translation, and outperforms strong LRMs including OpenAI-o1 and DeepSeeK-R1. Furthermore, we extend our method to the multilingual settings with 11 languages. With a carefully designed lightweight reward modeling in RL, we can simply transfer the strong MT ability from a single direction into multiple (i.e., 90) translation directions and achieve impressive multilingual MT performance.
This paper presents an end-to-end suite for multilingual information extraction and processing from image-based documents. The system uses Optical Character Recognition (Tesseract) to extract text in languages such as English, Hindi, and Tamil, and then a pipeline involving large language model APIs (Gemini) for cross-lingual translation, abstractive summarization, and re-translation into a target language. Additional modules add sentiment analysis (TensorFlow), topic classification (Transformers), and date extraction (Regex) for better document comprehension. Made available in an accessible Gradio interface, the current research shows a real-world application of libraries, models, and APIs to close the language gap and enhance access to information in image media across different linguistic environments
Translation-based strategies for cross-lingual transfer XLT such as translate-train -- training on noisy target language data translated from the source language -- and translate-test -- evaluating on noisy source language data translated from the target language -- are competitive XLT baselines. In XLT for token classification tasks, however, these strategies include label projection, the challenging step of mapping the labels from each token in the original sentence to its counterpart(s) in the translation. Although word aligners (WAs) are commonly used for label projection, the low-level design decisions for applying them to translation-based XLT have not been systematically investigated. Moreover, recent marker-based methods, which project labeled spans by inserting tags around them before (or after) translation, claim to outperform WAs in label projection for XLT. In this work, we revisit WAs for label projection, systematically investigating the effects of low-level design decisions on token-level XLT: (i) the algorithm for projecting labels between (multi-)token spans, (ii) filtering strategies to reduce the number of noisily mapped labels, and (iii) the pre-tokenization of the translated sentences. We find that all of these substantially impact translation-based XLT performance and show that, with optimized choices, XLT with WA offers performance at least comparable to that of marker-based methods. We then introduce a new projection strategy that ensembles translate-train and translate-test predictions and demonstrate that it substantially outperforms the marker-based projection. Crucially, we show that our proposed ensembling also reduces sensitivity to low-level WA design choices, resulting in more robust XLT for token classification tasks.
Neural machine translation (NMT) systems typically employ maximum a posteriori (MAP) decoding to select the highest-scoring translation from the distribution mass. However, recent evidence highlights the inadequacy of MAP decoding, often resulting in low-quality or even pathological hypotheses -- the decoding objective is not aligned with real-world translation quality. This paper proposes calibrating hypothesis likelihoods with translation quality from a distribution view by directly optimizing their Pearson correlation -- thereby enhancing the effectiveness of translation decoding. With our method, translation on large language models (LLMs) improves substantially after limited training (2K instances per direction). This improvement is orthogonal to those achieved through supervised fine-tuning, leading to substantial gains across a broad range of metrics and human evaluations -- even when applied to top-performing translation-specialized LLMs fine-tuned on high-quality translation data, such as Tower, or when compared to recent preference optimization methods, like CPO. Moreover, the calibrated translation likelihood can directly serve as a strong proxy for translation quality, closely approximating or even surpassing some state-of-the-art translation quality estimation models, like CometKiwi. Lastly, our in-depth analysis demonstrates that calibration enhances the effectiveness of MAP decoding, thereby enabling greater efficiency in real-world deployment. The resulting state-of-the-art translation model, which covers 10 languages, along with the accompanying code and human evaluation data, has been released to the community: https://github.com/moore3930/calibrating-llm-mt.
Current machine translation models provide us with high-quality outputs in most scenarios. However, they still face some specific problems, such as detecting which entities should not be changed during translation. In this paper, we explore the abilities of popular NMT models, including models from the OPUS project, Google Translate, MADLAD, and EuroLLM, to preserve entities such as URL addresses, IBAN numbers, or emails when producing translations between four languages: English, German, Polish, and Ukrainian. We investigate the quality of popular NMT models in terms of accuracy, discuss errors made by the models, and examine the reasons for errors. Our analysis highlights specific categories, such as emojis, that pose significant challenges for many models considered. In addition to the analysis, we propose a new multilingual synthetic dataset of 36,000 sentences that can help assess the quality of entity transfer across nine categories and four aforementioned languages.
The impact of Large Language Models (LLMs) has extended into literary domains. However, existing evaluation metrics prioritize mechanical accuracy over artistic expression and tend to overrate machine translation (MT) as being superior to experienced professional human translation. In the long run, this bias could result in a permanent decline in translation quality and cultural authenticity. In response to the urgent need for a specialized literary evaluation metric, we introduce LiTransProQA, a novel, reference-free, LLM-based question-answering framework designed specifically for literary translation evaluation. LiTransProQA uniquely integrates insights from professional literary translators and researchers, focusing on critical elements in literary quality assessment such as literary devices, cultural understanding, and authorial voice. Our extensive evaluation shows that while literary-finetuned XCOMET-XL yields marginal gains, LiTransProQA substantially outperforms current metrics, achieving up to 0.07 gain in correlation (ACC-EQ and Kendall's tau) and surpassing the best state-of-the-art metrics by over 15 points in adequacy assessments. Incorporating professional translator insights as weights further improves performance, highlighting the value of translator inputs. Notably, LiTransProQA approaches human-level evaluation performance comparable to trained linguistic annotators. It demonstrates broad applicability to open-source models such as LLaMA3.3-70b and Qwen2.5-32b, indicating its potential as an accessible and training-free literary evaluation metric and a valuable tool for evaluating texts that require local processing due to copyright or ethical considerations.
This paper explores the idea of using phonemes as a textual representation within a conventional multilingual simultaneous speech-to-speech translation pipeline, as opposed to the traditional reliance on text-based language representations. To investigate this, we trained an open-source sequence-to-sequence model on the WMT17 dataset in two formats: one using standard textual representation and the other employing phonemic representation. The performance of both approaches was assessed using the BLEU metric. Our findings shows that the phonemic approach provides comparable quality but offers several advantages, including lower resource requirements or better suitability for low-resource languages.
This paper introduces DATETIME, a new high-quality benchmark designed to evaluate the translation and reasoning abilities of a Large Language Model (LLM) on datetimes. A datetime is simply a date and a time, for example '11th.february.2023 ,1:12:31'. Datetimes are an interesting domain because they are intuitive and straightforward for humans to process but present significant challenges for LLMs. At the time of writing, no publicly available benchmark exists for systematically evaluating LLMs on datetime processing. Our experiments show that state-of-the-art models exhibit significant difficulty with tasks involving reasoning on datetimes, and that General Artificial Intelligence is still a distant aspiration. We hypothesize that working with datetimes necessitates translation and/or computation capabilities, and the tasks of the benchmark are organized accordingly. Significant dispersion in performance across models is observed with surprisingly poor performance even on apparently trivial tasks. Whilst frontier models such as ChatGPT, Claude and Llama3.1 have evidently been built and trained with datetime reasoning abilities, significant improvement is required for the open-source models.