arXiv subjects

André F. T. Martins

Publications and source records attributed to André F. T. Martins.

At least 19 recordsLinked to original sources

arXiv2026

Ouvia: A User-centered Framework for Measuring Usability of Speech Translation in Real-World Communication Scenarios

Speech translation (ST) is increasingly adopted in user applications, yet its evaluation largely focuses on decontextualized testbeds and holistic quality, rather than end users' communication needs. We introduce Ouvia, an evaluation framework for measuring user-perceived usability of speech translation outputs in real-world settings. Ouvia focuses on one-to-one communication: an English speaker needs to convey a request to a Portuguese speaker, and the message is automatically translated. Through a custom web app and multi-phase study design, we collect more than 1,750 such interactions in healthcare and everyday situations, mediated by four ST systems, involving speakers from three English dialects and two genders. We find that modern ST serves people only to a limited extent---only around half of interactions are rated as usable---with significant gaps in reported usability across demographic groups. Moreover, among quality metrics, we find that QA-based evaluation is a substantially stronger predictor of real-world usability than standard approaches. Together, these findings stress the importance of situated, user-centered evaluation frameworks that go beyond holistic quality scores and attend to who the technology serves---and how well.

cs.CL↗

arXiv2026

Length Generalization Needs Proper Regularization

Length generalization is the ability of sequential models to perform well on context lengths unseen during training. In this work, we show that the challenge of achieving length generalization is related not only to architectural choices such as positional encoding and the attention mechanism but also to the training procedure itself. We study how regularization affects length generalization and find that weight decay can hinder extrapolation. In contrast, dropout improves extrapolation when its placement within the architecture is reconsidered. In that regard, we show that the standard placement of dropout before layer normalization introduces a systematic distributional mismatch, and that applying dropout just before the linear projection resolves this issue. For example, a modified SmolLM3 with sliding window attention, continually pre-trained with dropout, can extrapolate perfectly to 64$\times$ on Needle-in-a-Haystack and far beyond the pre-training context size on RULER and HELMET. Mamba2 also benefits from dropout, suggesting an architecture-agnostic nature of the problem. We further propose Variance-Preserving Affine Dropout (VPAD), a new dropout strategy that substantially reduces the resulting pre-activation variance mismatch, leading to further extrapolation improvements in transformers.

cs.LG↗

arXiv2026

IndicQE-APE: A Consolidated Benchmark for Quality Estimation and Automatic Post-Editing for Indic Languages

Indic quality estimation (QE) and automatic post-editing (APE) data is spread across separate releases, so no single resource supports training and evaluation across tasks and language pairs on one footing. We consolidate the WMT 2020-2024 shared-task lineage with an extended English-Malayalam resource into IndicQE-APE: $126{,}754$ instances over nine directional pairs, with up to four label types aligned on the same segment, a direct assessment, a human post-edit, word-level tags and an error explanation, and a test set stratified over four difficulty axes. We benchmark six prompted LLMs and three COMET metrics on segment-level QE, and three systems on APE. Two of the axes are defined partly on direct assessment and select a compressed slice of it. Segments whose segment-level and token-level signals disagree are ranked below equally scored segments of the same language. Four-shot prompting costs every model at or below $3.4$B both correlation and output-format compliance. Unedited MT beats every APE system we run on three of the four pairs. The benchmark (https://huggingface.co/datasets/surrey-nlp/IndicQE-APE) and code (https://github.com/surrey-nlp/IndicQE-APE) are released.

cs.CL↗

arXiv2026

A Recipe for Long-Context Reasoning in Large Language Models via On-Policy Optimization and Distillation

Existing approaches to post-train models for long-context tasks face complementary limitations: (i) supervised fine-tuning (SFT) provides stable supervision but suffers from exposure bias; (ii) reinforcement learning methods such as Group Relative Policy Optimization (GRPO) train on model-generated trajectories but struggle with long-horizon credit assignment and sparse rewards; and (iii) on-policy distillation (OPD) provides dense token-level guidance but does not directly optimize task rewards. We study these complementary strategies for long-context alignment and derive a recipe that combines GRPO with OPD-style teacher guidance: the student learns from its own rollouts using outcome-level rewards, while a stronger teacher provides dense token-level regularization in place of the standard reference policy. This is especially useful when process-level supervision is difficult to obtain. To support this study, we introduce LongBlocks, a synthetic multilingual dataset spanning multi-hop reasoning, contextual grounding, and long-form generation. Through controlled ablations, we isolate the roles of cold-start initialization, teacher anchoring, and data mixing, showing that our recipe yields a more stable and effective path to long-context reasoning than GRPO or OPD while preserving short-context capabilities.

cs.CL↗

arXiv2026

MultiSynt/MT: Trillion-Token Multi-Parallel Pre-Training Data Translated Across 36 Languages

Open web-scale pre-training corpora remain concentrated in English, limiting multilingual LLM development. We introduce MultiSynt/MT, an open synthetic parallel corpus with approximately 4.8 trillion target-language tokens across 36 languages, produced by translating 100 billion high-quality Nemotron-CC tokens with Tower+ and OPUS-MT/HPLT-MT systems. For many medium- and lower-resource European languages, this is the largest openly available pre-training resource. Across five high- and medium-resource languages, reference LLMs trained on MultiSynt/MT reach the final score of HPLT 2.0, a native-data baseline, using roughly 72% fewer pre-training tokens, and outperform it by approximately 15% relative at a matched 100B-token training budget. Our analyses also identify evaluation blind spots: standard multiple-choice benchmarks miss translation-quality differences that a fluency-sensitive LLM-as-judge protocol recovers on the trained LLMs without detecting a deficit relative to its native-data baseline, while Norwegian idiomatic and culturally grounded tasks remain better served by native data. We release the corpus, including row-aligned translations from multiple systems, to support controlled research on multilingual pre-training data and evaluation.

cs.CL↗

arXiv2026

Self-Preference Bias in Rubric-Based Evaluation of Large Language Models

LLM-as-a-judge has become the de facto approach for evaluating LLM outputs. However, judges are known to exhibit self-preference bias (SPB): they tend to favor outputs produced by themselves or by models from their own family. This skews evaluations and, thus, hinders model development, especially in settings of recursive self-improvement. We present the first study of SPB in rubric-based evaluation, an increasingly popular benchmarking paradigm where judges issue binary verdicts on individual evaluation criteria, instead of assigning holistic scores or rankings. Using IFEval and LiveCodeBench, benchmarks with programmatically verifiable rubrics, we show that SPB persists even when evaluation criteria are entirely objective: among rubrics where generators fail, judges can be more than 50% more likely to incorrectly mark them as satisfied when the output is their own. We also find that, similarly to other evaluation paradigms, ensembling multiple judges helps mitigate SPB, but without fully eliminating it. On HealthBench, a medical chat benchmark with subjective rubrics, we observe that SPB skews model scores by up to 10 points, a potentially decisive margin when ranking frontier models. We analyze the factors that drive SPB in this setting, finding that negative rubrics and subjective topics like communication and emergency referrals are particularly susceptible.

cs.CL↗

arXiv2026

What's Holding Back Latent Visual Reasoning?

Humans can approach complex visual problems by mentally simulating intermediate visual steps, rather than reasoning through language alone. Inspired by this, several works on Vision-Language Models have recently explored chain-of-thought reasoning with continuous latent tokens as intermediate visual imagination steps. In this work, we investigate how recent models leverage such latent tokens. Surprisingly, we find that model accuracy is unaffected when latent tokens are replaced by uninformative dummy tokens. This indicates that latent tokens play a minimal causal role in the model's final prediction. To better understand this phenomenon, we analyze both the training signal provided by oracle latent representations and the quality of the latent tokens generated at inference time. Our experiments reveal two crucial issues holding back latent visual reasoning: First, in most existing datasets, oracle latent tokens provide limited additional information beyond the original image and do not substantially simplify the task, leading models to ignore them during training and effectively bypassing them at inference time. When fine-tuned on a diagnostic dataset, in which latent tokens provide sufficient support for the final prediction, we show that models can causally rely on them. Second, the latent tokens produced at inference time deviate from their corresponding oracle representations, collapsing to a narrow region and preventing benefits even when the model relies on them. Overall, our findings suggest that future progress in latent visual reasoning depends on two key pillars: high-quality datasets with informative intermediate steps and more precise latent token prediction.

cs.CV↗

arXiv2026

DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention

Current hierarchical attention methods, such as NSA and InfLLMv2, select the top-k relevant key-value (KV) blocks based on coarse attention scores and subsequently apply fine-grained softmax attention on the selected tokens. However, the top-k operation assumes the number of relevant tokens for any query is fixed and it precludes the gradient flow between the sparse and dense stages. In this work, we propose DashAttention (Differentiable and Adaptive Sparse Hierarchical Attention), which leverages the adaptively sparse $α$-entmax transformation to select a variable number of blocks according to the current query in the first stage. This in turn provides a prior for the second-stage softmax attention, keeping the entire hierarchy fully differentiable. Contrary to other hierarchical attention methods, we show that DashAttention is non-dispersive, translating to better long-context modeling ability. Experiments with large language models (LLMs) show that DashAttention achieves comparable accuracy as full attention with 75% sparsity and a better Pareto frontier than NSA and InfLLMv2, especially in high-sparsity regimes. We also provide an efficient, GPU-aware implementation of DashAttention in Triton, which achieves a speedup of up to over FlashAttention-3 at inference time. Overall, DashAttention offers a cost-effective strategy to model long contexts.

cs.CL↗

arXiv2026

SEQUOR: A Multi-Turn Benchmark for Realistic Constraint Following

In a conversation, a helpful assistant must reliably follow user directives, even as they refine, modify, or contradict earlier requests. Yet most instruction-following benchmarks focus on single-turn or short multi-turn scenarios, leaving open how well models handle long-horizon instruction-following tasks. To bridge this gap, we present SEQUOR, an automatic benchmark for evaluating constraint adherence in long multi-turn conversations. SEQUOR consists of simulated persona-driven interactions built with constraints extracted from real-world conversations. Our results show that even when following a single constraint, instruction-following accuracy consistently decreases as the conversation grows longer, with drops exceeding 11%. This decline becomes larger when models have to follow multiple constraints simultaneously, reducing their accuracy by over 40%. In scenarios where constraints are added or replaced at arbitrary points of the conversation, model accuracy decreases by more than 9%. Taken together, our results reveal that current models still struggle to follow user instructions in multi-turn conversations, and provide a way for better measuring instruction-following capabilities in assistants.

cs.CL↗

arXiv2026

Should We Still Pretrain Encoders with Masked Language Modeling?

Learning high-quality text representations is fundamental to a wide range of NLP tasks. While encoder pretraining has traditionally relied on Masked Language Modeling (MLM), recent evidence suggests that decoder models pretrained with Causal Language Modeling (CLM) can be effectively repurposed as encoders, often surpassing traditional encoders on text representation benchmarks. However, it remains unclear whether these gains reflect an inherent advantage of the CLM objective or arise from confounding factors such as model and data scale. In this paper, we address this question through a series of large-scale, carefully controlled pretraining ablations, training a total of 38 models ranging from 210 million to 1 billion parameters, and conducting over 15,000 fine-tuning and evaluation runs. We find that while training with MLM generally yields better performance across text representation tasks, CLM-trained models are more data-efficient and demonstrate improved fine-tuning stability. Building on these findings, we experimentally show that a biphasic training strategy that sequentially applies CLM and then MLM, achieves optimal performance under a fixed computational training budget. Moreover, we demonstrate that this strategy becomes more appealing when initializing from readily available pretrained CLM models, reducing the computational burden needed to train best-in-class encoder models. We release all project artifacts at https://hf.co/MLMvsCLM to foster further research.

cs.CL↗

arXiv2026

MATH-PT: A Math Reasoning Benchmark for European and Brazilian Portuguese

The use of large language models (LLMs) for complex mathematical reasoning is an emergent area of research, with fast progress in methods, models, and benchmark datasets. However, most mathematical reasoning evaluations exhibit a significant linguistic bias, with the vast majority of benchmark datasets being exclusively in English or (at best) translated from English. We address this limitation by introducing {\sc Math-PT}, a novel dataset comprising 1,729 mathematical problems written in European and Brazilian Portuguese. {\sc Math-PT} is curated from a variety of high-quality native sources, including mathematical Olympiads, competitions, and exams from Portugal and Brazil. We present a comprehensive benchmark of current state-of-the-art LLMs on {\sc Math-PT}, revealing that frontier reasoning models achieve strong performance in multiple choice questions compared to open weight models, but that their performance decreases for questions with figures or open-ended questions. To facilitate future research, we release the benchmark dataset and model outputs.

cs.CL↗

arXiv2026

Long-Context Generalization with Sparse Attention

Transformer-based architectures traditionally employ softmax to compute attention weights, which produces dense distributions over all tokens in a sequence. While effective in many settings, this density has been shown to be detrimental for tasks that demand precise focus on fixed-size patterns: as sequence length increases, non-informative tokens accumulate attention probability mass, leading to dispersion and representational collapse. We show in this paper that dynamically sparse attention mechanisms using $α$-entmax can avoid these issues, due to their ability to assign exact zeros to irrelevant tokens. Furthermore, we introduce Adaptive-Scalable Entmax (ASEntmax), which endows $α$-entmax with a learnable temperature parameter, allowing the attention distribution to interpolate between sparse (pattern-focused) and dense (softmax-like) regimes. Our empirical evaluation on synthetic tasks and language modeling demonstrates that ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature $α$-entmax baselines, achieving up to 1000$\times$ length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance, including better perplexity trends and higher retrieval accuracies at 8$\times$ training length.

cs.CL↗

arXiv2026

EuroLLM-22B: Technical Report

This report presents EuroLLM-22B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-22B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. Across a broad set of multilingual benchmarks, EuroLLM-22B demonstrates strong performance in reasoning, instruction following, and translation, achieving results competitive with models of comparable size. To support future research, we release our base and instruction-tuned models, our multilingual web pretraining data and updated EuroBlocks instruction datasets, as well as our pre-training and evaluation codebases.

cs.CL↗

arXiv2026

Discrete Latent Structure in Neural Networks

Many types of data from fields including natural language processing, computer vision, and bioinformatics, are well represented by discrete, compositional structures such as trees, sequences, or matchings. Latent structure models are a powerful tool for learning to extract such representations, offering a way to incorporate structural bias, discover insight about the data, and interpret decisions. However, effective training is challenging, as neural networks are typically designed for continuous computation. This text explores three broad strategies for learning with discrete latent structure: continuous relaxation, surrogate gradients, and probabilistic estimation. Our presentation relies on consistent notations for a wide range of models. As such, we reveal many new connections between latent structure learning strategies, showing how most consist of the same small set of fundamental building blocks, but use them differently, leading to substantially different applicability and properties.

cs.LG↗

arXiv2025

Learning with the $p$-adics

Existing machine learning frameworks operate over the field of real numbers ($\mathbb{R}$) and learn representations in real (Euclidean or Hilbert) vector spaces (e.g., $\mathbb{R}^d$). Their underlying geometric properties align well with intuitive concepts such as linear separability, minimum enclosing balls, and subspace projection; and basic calculus provides a toolbox for learning through gradient-based optimization. But is this the only possible choice? In this paper, we study the suitability of a radically different field as an alternative to $\mathbb{R}$ -- the ultrametric and non-archimedean space of $p$-adic numbers, $\mathbb{Q}_p$. The hierarchical structure of the $p$-adics and their interpretation as infinite strings make them an appealing tool for code theory and hierarchical representation learning. Our exploratory theoretical work establishes the building blocks for classification, regression, and representation learning with the $p$-adics, providing learning models and algorithms. We illustrate how simple Quillian semantic networks can be represented as a compact $p$-adic linear network, a construction which is not possible with the field of reals. We finish by discussing open problems and opportunities for future research enabled by this new framework.

cs.LG↗

arXiv2025

Fine-Grained Reward Optimization for Machine Translation using Error Severity Mappings

Reinforcement learning (RL) has been proven to be an effective and robust method for training neural machine translation systems, especially when paired with powerful reward models that accurately assess translation quality. However, most research has focused on RL methods that use sentence-level feedback, leading to inefficient learning signals due to the reward sparsity problem -- the model receives a single score for the entire sentence. To address this, we propose a novel approach that leverages fine-grained, token-level quality assessments along with error severity levels using RL methods. Specifically, we use xCOMET, a state-of-the-art quality estimation system, as our token-level reward model. We conduct experiments on small and large translation datasets with standard encoder-decoder and large language models-based machine translation systems, comparing the impact of sentence-level versus fine-grained reward signals on translation quality. Our results show that training with token-level rewards improves translation quality across language pairs over baselines according to both automatic and human evaluation. Furthermore, token-level reward optimization improves training stability, evidenced by a steady increase in mean rewards over training epochs.

cs.CL↗

arXiv2025

TowerVision: Understanding and Improving Multilinguality in Vision-Language Models

Despite significant advances in vision-language models (VLMs), most existing work follows an English-centric design process, limiting their effectiveness in multilingual settings. In this work, we provide a comprehensive empirical study analyzing the impact of several multilingual design choices, such as training data composition, encoder selection, and text backbones. The result is TowerVision, a family of open multilingual VLMs for both image-text and video-text tasks, built upon the multilingual text-only model Tower+. TowerVision achieves competitive performance on multiple multimodal multilingual benchmarks and shows particular strength in culturally grounded tasks and multimodal translation. By incorporating visual and cultural context during fine-tuning, our models surpass existing approaches trained on substantially larger datasets, as demonstrated on ALM-Bench and Multi30K (image tasks) and ViMUL-Bench (video tasks). Alongside the models, we release VisionBlocks, a high-quality, curated vision-language dataset. Our findings highlight that multilingual vision-language training data substantially improves cross-lingual generalization -- both from high-resource to underrepresented languages and vice versa -- and that instruction-tuned LLMs are not always the optimal initialization point. To support further research, we publicly release all models, data, and training recipes.

cs.LG↗

arXiv2025

Zero-shot Benchmarking: A Framework for Flexible and Scalable Automatic Evaluation of Language Models

As language models improve and become capable of performing more complex tasks across modalities, evaluating them automatically becomes increasingly challenging. Developing strong and robust task-specific automatic metrics gets harder, and human-annotated test sets -- which are expensive to create -- saturate more quickly. A compelling alternative is to design reliable strategies to automate the creation of test data and evaluation, but previous attempts either rely on pre-existing data, or focus solely on individual tasks. We present Zero-shot Benchmarking (ZSB), a framework for creating high-quality benchmarks for any task by leveraging language models for both synthetic test data creation and evaluation. ZSB is simple and flexible: it requires only the creation of a prompt for data generation and one for evaluation; it is scalable to tasks and languages where collecting real-world data is costly or impractical; it is model-agnostic, allowing the creation of increasingly challenging benchmarks as models improve. To assess the effectiveness of our framework, we create benchmarks for five text-only tasks and a multi-modal one: general capabilities in four languages (English, Chinese, French, and Korean), translation, and general vision-language capabilities in English. We then rank a broad range of open and closed systems on our benchmarks. ZSB rankings consistently correlate strongly with human rankings, outperforming widely-adopted standard benchmarks. Through ablations, we find that strong benchmarks can be created with open models, and that judge model size and dataset variety are crucial drivers of performance. We release all our benchmarks, and code to reproduce our experiments and to produce new benchmarks.

cs.CL↗