arXiv:2511.10780v1 Announce Type: new 
Abstract: In this paper, we introduce TEDxTN, the first publicly available Tunisian Arabic to English speech translation dataset. This work is in line with the ongoing effort to mitigate the data scarcity obstacle for a number of Arabic dialects. We collected, segmented, transcribed and translated 108 TEDx talks following our internally developed annotations guidelines. The collected talks represent 25 hours of speech with code-switching that cover speakers with various accents from over 11 different regions of Tunisia. We make the annotation guidelines and corpus publicly available. This will enable the extension of TEDxTN to new talks as they become available. We also report results for strong baseline systems of Speech Recognition and Speech Translation using multiple pre-trained and fine-tuned end-to-end models. This corpus is the first open source and publicly available speech translation corpus of Code-Switching Tunisian dialect. We believe that this is a valuable resource that can motivate and facilitate further research on the natural language processing of Tunisian Dialect.

يقدم مشروع TEDxTN أول مجموعة بيانات متاحة للجمهور لترجمة الكلام من العربية التونسية إلى الإنجليزية. تتضمن هذه المجموعة 108 محاضرات من TEDx، بإجمالي 25 ساعة من الكلام، مع تغييرات في اللغة وأصوات متنوعة من مختلف المناطق في تونس. يهدف هذا المشروع إلى معالجة مشكلة نقص البيانات للعديد من اللهجات العربية، ويأتي مع إرشادات توضيحية متاحة للجمهور، مما يتيح التوسع في المستقبل.

El proyecto TEDxTN presenta el primer conjunto de datos de traducción de discurso disponible públicamente para el árabe tunecino al inglés. Este conjunto incluye 108 charlas TEDx, totalizando 25 horas de discurso, con cambios de código y diversos acentos regionales de Túnez. El corpus tiene como objetivo abordar el problema de la escasez de datos para los dialectos árabes y se acompaña de pautas de anotación disponibles públicamente, lo que permite futuras expansiones.

Le projet TEDxTN présente le premier ensemble de données de traduction de discours disponible publiquement pour l'arabe tunisien vers l'anglais. Cet ensemble comprend 108 discours TEDx, totalisant 25 heures de discours, avec des changements de code et divers accents régionaux de Tunisie. Le corpus vise à résoudre le problème de la rareté des données pour les dialectes arabes et est accompagné de directives d'annotation disponibles publiquement, permettant des extensions futures.

The TEDxTN project introduces the first publicly available speech translation dataset for Tunisian Arabic to English. This dataset includes 108 TEDx talks, totaling 25 hours of speech, featuring code-switching and various regional accents from Tunisia. The corpus aims to address the data scarcity issue for Arabic dialects and is accompanied by publicly available annotation guidelines, enabling future expansions.

TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English

<A HREF="https://www.businessinsider.com/meta-vibes-ai-internal-documents-show-daily-active-users-2025-11"><IMG VSPACE="4" HSPACE="4" BORDER="0" ALIGN="RIGHT" SRC="http://www.techmeme.com/251118/i40.jpg"></A>
<A HREF="http://www.techmeme.com/251118/p40#a251118p40" TITLE="Techmeme permalink"><IMG WIDTH=11 HEIGHT=12 SRC="http://www.techmeme.com/img/pml.png" STYLE="border:none;padding:0;margin:0;"></A> Pranav Dixit / <A HREF="https://www.businessinsider.com/">Business Insider</A>: 
<A HREF="https://www.businessinsider.com/meta-vibes-ai-internal-documents-show-daily-active-users-2025-11">Internal docs show Meta's Vibes AI video feed has about 2M DAUs as of November 9, up 1% from the previous week; ~52% of returning users prompted the AI</A>&nbsp; &mdash;&nbsp; - Internal documents show how Meta's new Vibes AI video feed is performing across its biggest markets.

تكشف الوثائق الداخلية أن خدمة الفيديو Vibes AI التابعة لشركة ميتا تضم حوالي 2 مليون مستخدم نشط يوميًا اعتبارًا من 9 نوفمبر، بزيادة قدرها 1% عن الأسبوع السابق. وقد تفاعل حوالي 52% من المستخدمين العائدين مع ميزة الذكاء الاصطناعي. تسلط هذه البيانات الضوء على زيادة التفاعل مع المحتوى المدعوم بالذكاء الاصطناعي من ميتا، مما يشير إلى اتجاه إيجابي في تفاعل المستخدمين واحتفاظهم.

Documentos internos revelan que el feed de video Vibes AI de Meta tiene aproximadamente 2 millones de usuarios activos diarios (UAD) a partir del 9 de noviembre, lo que representa un aumento del 1% respecto a la semana anterior. Aproximadamente el 52% de los usuarios que regresan interactuaron con la función de IA. Estos datos de rendimiento destacan el creciente compromiso con el contenido impulsado por IA de Meta, indicando una tendencia positiva en la interacción y retención de usuarios.

Des documents internes révèlent que le flux vidéo Vibes AI de Meta compte environ 2 millions d'utilisateurs actifs quotidiens (UAQ) au 9 novembre, soit une augmentation de 1 % par rapport à la semaine précédente. Environ 52 % des utilisateurs revenants ont interagi avec la fonctionnalité AI. Ces données de performance mettent en lumière l'engagement croissant envers le contenu alimenté par l'IA de Meta, indiquant une tendance positive dans l'interaction et la fidélisation des utilisateurs.

Internal documents reveal that Meta's Vibes AI video feed has approximately 2 million daily active users (DAUs) as of November 9, reflecting a 1% increase from the previous week. Notably, around 52% of returning users interacted with the AI feature. This performance data highlights the growing engagement with Meta's AI-driven content, indicating a positive trend in user interaction and retention.

Internal docs show Meta's Vibes AI video feed has about 2M DAUs as of November 9, up 1% from the previous week; ~52% of returning users prompted the AI (Pranav Dixit/Business Insider)

GlobalFoundries acquires AMF to become top silicon photonics foundry.
The post <a href="https://www.eetimes.com/gf-targets-1-billion-silicon-photonics-revenue-with-amf-acquisition/">GF Targets $1 Billion Silicon Photonics Revenue with AMF Acquisition</a> appeared first on <a href="https://www.eetimes.com">EE Times</a>.

استحوذت شركة GlobalFoundries على AMF، مما يضعها في موقع رائد في سوق الفوتونيك السيليكون. تهدف هذه الصفقة إلى تحقيق إيرادات تصل إلى مليار دولار من الفوتونيك السيليكون، وهي تقنية تدمج المكونات الضوئية مع الدوائر السيليكونية، مما يعزز سرعة وكفاءة نقل البيانات. تعكس هذه الخطوة التزام GlobalFoundries بتوسيع قدراتها في التقنيات المتقدمة وتلبية الطلب المتزايد على حلول الحوسبة عالية الأداء.

GlobalFoundries ha adquirido AMF, posicionándose como un jugador líder en el mercado de la fotónica de silicio. Esta adquisición tiene como objetivo generar 1.000 millones de dólares en ingresos de fotónica de silicio, una tecnología que integra componentes ópticos con circuitos de silicio, mejorando la velocidad y eficiencia de la transmisión de datos. Este movimiento refleja el compromiso de GlobalFoundries de expandir sus capacidades en tecnologías avanzadas y satisfacer la creciente demanda de soluciones informáticas de alto rendimiento.

GlobalFoundries a acquis AMF, se positionnant comme un acteur majeur sur le marché de la photonique silicium. Cette acquisition vise à générer 1 milliard de dollars de revenus provenant de la photonique silicium, une technologie qui intègre des composants optiques avec des circuits en silicium, améliorant ainsi la vitesse et l'efficacité de la transmission des données. Ce mouvement reflète l'engagement de GlobalFoundries à élargir ses capacités dans les technologies avancées et à répondre à la demande croissante de solutions informatiques haute performance.

GlobalFoundries has acquired AMF, positioning itself as a leading player in the silicon photonics market. This acquisition aims to generate $1 billion in revenue from silicon photonics, a technology that integrates optical components with silicon circuits, enhancing data transmission speeds and efficiency. The move reflects GlobalFoundries' commitment to expanding its capabilities in advanced technologies and meeting the growing demand for high-performance computing solutions.

GF Targets $1 Billion Silicon Photonics Revenue with AMF Acquisition

Black Friday is just two weeks away, and in the lead-up to the sales event, I've collected the best early Chromebook deals across major retailers.

يقترب يوم الجمعة السوداء بعد أسبوعين، وقد بدأت العروض المبكرة على أجهزة Chromebook تظهر في المتاجر الكبرى. يبرز هذا المقال أكثر من 20 عرضًا مبكرًا، مما يوفر للمستهلكين فرصة للتوفير في مشترياتهم خلال العطلات.

El Black Friday se acerca en dos semanas, y ya han comenzado a aparecer ofertas anticipadas en Chromebooks en los principales minoristas. Este artículo destaca más de 20 ventas anticipadas, brindando a los consumidores la oportunidad de ahorrar en sus compras navideñas.

Le Black Friday approche dans deux semaines, et des offres précoces sur les Chromebooks commencent à apparaître chez les principaux détaillants. Cet article met en avant plus de 20 ventes anticipées, offrant aux consommateurs l'occasion d'économiser sur leurs achats de vacances.

Black Friday is approaching in two weeks, and early Chromebook deals have started to appear across major retailers. This article highlights over 20 early sales, providing consumers with an opportunity to save on their holiday shopping.

Best early Black Friday Chromebook deals 2025: 20+ sales out early

Buy now, pay later firm says pay has risen by 60% with staff numbers mostly cut by natural attrition and tech investmentKlarna has claimed that AI-related savings have allowed the buy now, pay later company to increase staff salaries by nearly 60%, but hinted it could slash more jobs after nearly halving its workforce over the past three years.Chief executive Sebastian Siemiatkowski said headcount had dropped from 5,527 to 2,907 since 2022, mostly as a result of natural attrition, with departing staff replaced by technology rather than by new staff members. <a href="https://www.theguardian.com/business/2025/nov/18/buy-now-pay-later-klarna-ai-helped-halve-staff-boost-pay">Continue reading...</a>

أعلنت شركة كلارنا، المتخصصة في الدفع لاحقًا، عن زيادة بنسبة 60% في رواتب موظفيها، مشيرةً إلى أن هذه الزيادة جاءت نتيجة التوفير الناتج عن الاستثمارات في الذكاء الاصطناعي. وقد خفضت الشركة عدد موظفيها من 5,527 إلى 2,907 خلال السنوات الثلاث الماضية، وذلك بشكل رئيسي من خلال الاستقالة الطبيعية، حيث تم استبدال الموظفين المغادرين بالتكنولوجيا بدلاً من توظيف موظفين جدد. وأشار الرئيس التنفيذي سيباستيان سييمياتكوفسكي إلى أنه قد يتم تقليص المزيد من الوظائف مع استمرار الشركة في الاستفادة من التكنولوجيا لتعزيز الكفاءة.

Klarna, una empresa de 'compra ahora, paga después', ha informado de un aumento del 60% en los salarios de su personal, atribuyendo este incremento a los ahorros generados por inversiones en inteligencia artificial. La compañía ha reducido su plantilla de 5,527 a 2,907 en los últimos tres años, principalmente por medio de la baja natural, con la tecnología reemplazando a los empleados que se van en lugar de contratar nuevos. El CEO Sebastian Siemiatkowski indicó que podrían producirse más despidos a medida que la empresa continúe aprovechando la tecnología para mejorar la eficiencia.

Klarna, une entreprise de paiement différé, a annoncé une augmentation de 60 % des salaires de ses employés, attribuant cette hausse aux économies réalisées grâce aux investissements en IA. L'entreprise a réduit son effectif de 5 527 à 2 907 au cours des trois dernières années, principalement par attrition naturelle, la technologie remplaçant les employés partants au lieu d'embaucher de nouveaux. Le PDG Sebastian Siemiatkowski a indiqué que d'autres suppressions d'emplois pourraient se produire alors que l'entreprise continue d'exploiter la technologie pour améliorer son efficacité.

Klarna, a buy now, pay later company, has reported a 60% increase in staff salaries, attributing this rise to savings generated from AI investments. The company has reduced its workforce from 5,527 to 2,907 over the past three years, primarily through natural attrition, with technology replacing departing employees instead of hiring new staff. CEO Sebastian Siemiatkowski indicated that further job cuts could occur as the company continues to leverage technology to enhance efficiency.

Klarna says AI drive has helped halve staff numbers and boost pay

Evaluating Generative AI: A Novel Metric - Perceptual Diversity

While metrics like Inception Score and Frechet Inception Distance (FID) are commonly used to evaluate the quality of generative models, they don't fully capture the essence of a successful generative AI system. Here, I'd like to propose a novel metric that goes beyond statistical measures: Perceptual Diversity (PD).

What is Perceptual Diversity?

Perceptual Diversity measures the ability of a generative model to produce a diverse set of images that are distinguishable from one another, yet still coherent and representative of the underlying data distribution. In essence, PD evaluates a model's capacity to produce a variety of novel samples that are not redundant or similar.

Example: Generative AI for Architectural Design

Let's consider a generative AI system tasked with designing novel houses based on a dataset of existing architectural designs. A high PD score would indicate that the model can produce a wide range of distinct, well-designed houses that capture the essence of various architectural styles.

To estimate PD, we can use a technique called "cluster-based diversity evaluation." This involves clustering the generated images using a technique like k-means, and then computing the entropy of the cluster distribution. The higher the entropy, the more diverse the generated samples.

Example Results

Using a Generative Adversarial Network (GAN) model trained on a dataset of 1000 architectural designs, we obtained the following results:

<ul>
<li>Average Inception Score: 5.2</li>
<li>Average FID Score: 10.5</li>
<li>Average Perceptual Diversity (PD): 0.85</li>
</ul>

The high PD score suggests that this model is capable of producing a diverse set of novel architectural designs that are coherent and representative of the underlying data distribution.

Conclusion

Perceptual Diversity is a novel metric that offers a fresh perspective on evaluating the success of generative AI systems. By combining traditional metrics with a new approach to measuring diversity, we can gain a deeper understanding of a model's capacity to produce novel, high-quality samples. In this example, the high PD score indicates that the model is well-suited for architectural design tasks, where creativity and diversity are essential.




Publicado automáticamente

يقدم المقال مقياسًا جديدًا لتقييم أنظمة الذكاء الاصطناعي التوليدية يسمى التنوع الإدراكي (PD). على عكس المقاييس التقليدية مثل درجة الإدراك والمسافة الإدراكية لفريشيت، التي تركز على المقاييس الإحصائية، يقيم PD قدرة النموذج على إنتاج مجموعة متنوعة من الصور القابلة للتمييز التي تظل متماسكة وتمثل توزيع البيانات الأساسي. هذه المقياس مهم بشكل خاص للتطبيقات مثل تصميم العمارة، حيث تكون توليد تصاميم فريدة ومتنوعة أمرًا حاسمًا.

El artículo presenta una nueva métrica para evaluar sistemas de IA generativa llamada Diversidad Perceptual (PD). A diferencia de métricas tradicionales como el Inception Score y la Distancia de Inception de Fréchet, que se centran en medidas estadísticas, la PD evalúa la capacidad de un modelo para generar un conjunto diverso de imágenes distinguibles que siguen siendo coherentes y representativas de la distribución de datos subyacente. Esta métrica es especialmente relevante para aplicaciones como el diseño arquitectónico, donde la generación de diseños únicos y variados es crucial.

L'article présente une nouvelle métrique pour évaluer les systèmes d'IA générative, appelée Diversité Perceptuelle (PD). Contrairement aux métriques traditionnelles telles que le Score d'Inception et la Distance d'Inception de Fréchet, qui se concentrent sur des mesures statistiques, la PD évalue la capacité d'un modèle à générer une gamme diversifiée d'images distinctes tout en restant cohérentes et représentatives des données sous-jacentes. Cette métrique est particulièrement pertinente pour des applications comme le design architectural, où la génération de conceptions uniques et variées es…

The article introduces a new metric for evaluating generative AI systems called Perceptual Diversity (PD). Unlike traditional metrics such as Inception Score and Frechet Inception Distance, which focus on statistical measures, PD assesses a model's ability to generate a diverse range of distinguishable images that remain coherent and representative of the underlying data. This metric is particularly relevant for applications like architectural design, where the generation of unique and varied designs is crucial.

**Evaluating Generative AI: A Novel Metric - Perceptual Dive

Black Friday is Nov. 28, but major retailers like Amazon, Best Buy, and Walmart are already slashing prices on top tech.

Best early Black Friday deals 2025: 55+ deals on TVs, laptops, streaming, and more

arXiv:2511.11334v1 Announce Type: new 
Abstract: The rapid advancement of large language models (LLMs) has not been matched by their evaluation in low-resource languages, especially Southeast Asian languages like Lao. To fill this gap, we introduce LaoBench, the first large-scale, high-quality, and multidimensional benchmark dataset dedicated to assessing LLMs' comprehensive language understanding and reasoning abilities in Lao. LaoBench comprises over 17,000 carefully curated samples spanning three core dimensions: knowledge application, K12 foundational education, and bilingual translation among Lao, Chinese, and English. The dataset is divided into open-source and closed-source subsets, with the closed-source portion enabling black-box evaluation on an official platform to ensure fairness and data security. Our data construction pipeline integrates expert human curation with automated agent-assisted verification, ensuring linguistic accuracy, cultural relevance, and educational value. Benchmarking multiple state-of-the-art LLMs on LaoBench reveals that current models still face significant challenges in mastering Lao across diverse tasks. We hope LaoBench will catalyze further research and development of AI technologies for underrepresented Southeast Asian languages.

لاوبينش هو مجموعة بيانات جديدة على نطاق واسع تهدف إلى تقييم نماذج اللغة الكبيرة (LLMs) باللغة اللاوسية. تتكون من أكثر من 17,000 عينة مختارة بعناية، وتقيّم تطبيق المعرفة، والتعليم الأساسي، والترجمة الثنائية بين اللاوسية والصينية والإنجليزية. تهدف هذه المجموعة إلى تحسين قدرات الفهم والتفكير لنماذج اللغة الكبيرة في اللغات ذات الموارد المحدودة، مما يعالج التحديات الحالية التي تواجه النماذج في إتقان اللغة اللاوسية.

LaoBench es un nuevo conjunto de datos de referencia a gran escala destinado a evaluar modelos de lenguaje grandes (LLMs) en el idioma lao. Con más de 17,000 muestras cuidadosamente seleccionadas, evalúa la aplicación del conocimiento, la educación básica y la traducción bilingüe entre lao, chino e inglés. Este conjunto de datos busca mejorar las capacidades de comprensión y razonamiento de los LLMs en idiomas de bajos recursos, abordando los desafíos actuales que enfrentan los modelos para dominar el lao.

LaoBench est un nouvel ensemble de données de référence à grande échelle destiné à évaluer les grands modèles de langage (LLMs) en lao. Il comprend plus de 17 000 échantillons soigneusement sélectionnés qui évaluent l'application des connaissances, l'éducation fondamentale et la traduction bilingue entre le lao, le chinois et l'anglais. Cet ensemble de données vise à améliorer les capacités de compréhension et de raisonnement des LLMs dans les langues à faibles ressources, en répondant aux défis actuels rencontrés par les modèles dans la maîtrise du lao.

LaoBench is a newly introduced large-scale benchmark dataset aimed at evaluating large language models (LLMs) in the Lao language. It consists of over 17,000 curated samples that assess knowledge application, foundational education, and bilingual translation among Lao, Chinese, and English. The dataset is designed to enhance the understanding and reasoning capabilities of LLMs in low-resource languages, addressing the current challenges faced by models in mastering Lao.

LaoBench: A Large-Scale Multidimensional Lao Benchmark for Large Language Models

arXiv:2511.11427v1 Announce Type: new 
Abstract: Referring Expression Comprehension (REC) requires models to localize objects in images based on natural language descriptions. Research on the area remains predominantly English-centric, despite increasing global deployment demands. This work addresses multilingual REC through two main contributions. First, we construct a unified multilingual dataset spanning 10 languages, by systematically expanding 12 existing English REC benchmarks through machine translation and context-based translation enhancement. The resulting dataset comprises approximately 8 million multilingual referring expressions across 177,620 images, with 336,882 annotated objects. Second, we introduce an attention-anchored neural architecture that uses multilingual SigLIP2 encoders. Our attention-based approach generates coarse spatial anchors from attention distributions, which are subsequently refined through learned residuals. Experimental evaluation demonstrates competitive performance on standard benchmarks, e.g. achieving 86.9% accuracy at IoU@50 on RefCOCO aggregate multilingual evaluation, compared to an English-only result of 91.3%. Multilingual evaluation shows consistent capabilities across languages, establishing the practical feasibility of multilingual visual grounding systems. The dataset and model are available at $\href{https://multilingual.franreno.com}{multilingual.franreno.com}$.

تتناول الدراسة حول فهم التعبيرات متعددة اللغات (REC) كيفية تحديد مواقع الأجسام في الصور بناءً على أوصاف باللغة الطبيعية. على الرغم من الطلب العالمي المتزايد على التطبيقات متعددة اللغات، فإن الأبحاث الحالية تركز بشكل أساسي على اللغة الإنجليزية. يقدم هذا العمل مجموعة بيانات متعددة اللغات موحدة تغطي 10 لغات، تم إنشاؤها من خلال توسيع 12 معيارًا إنجليزيًا عبر الترجمة الآلية، مما أسفر عن حوالي 8 ملايين تعبير متعدد اللغات عبر 177620 صورة و336882 كائنًا موصوفًا. بالإضافة إلى ذلك، تم تقديم بنية عصبية جديدة تعتمد على الانتباه لتحسين أداء REC.

El estudio sobre la Comprensión de Expresiones Referenciales (REC) se centra en la localización de objetos en imágenes a partir de descripciones en lenguaje natural. A pesar de la creciente demanda global de aplicaciones multilingües, la investigación existente ha sido predominantemente centrada en el inglés. Este trabajo presenta un conjunto de datos multilingüe unificado que abarca 10 idiomas, creado al expandir 12 benchmarks en inglés mediante traducción automática, resultando en aproximadamente 8 millones de expresiones en 177,620 imágenes y 336,882 objetos anotados. Además, se introduce u…

Cette étude sur la compréhension des expressions référentielles (REC) se concentre sur la localisation d'objets dans des images à l'aide de descriptions en langage naturel. Malgré le besoin mondial d'applications multilingues, les recherches existantes ont été principalement centrées sur l'anglais. Ce travail introduit un ensemble de données multilingue unifié couvrant 10 langues, créé en élargissant 12 benchmarks anglais par traduction automatique, aboutissant à environ 8 millions d'expressions à travers 177 620 images et 336 882 objets annotés. De plus, une nouvelle architecture neuronale an…

The study on Referring Expression Comprehension (REC) focuses on localizing objects in images using natural language descriptions. Despite the global need for multilingual applications, existing research has been primarily English-centric. This work introduces a unified multilingual dataset covering 10 languages, created by expanding 12 English benchmarks through machine translation, resulting in about 8 million expressions across 177,620 images and 336,882 annotated objects. Additionally, a new attention-anchored neural architecture is proposed to enhance REC performance.

Comprehension of Multilingual Expressions Referring to Target Objects in Visual Inputs

arXiv:2510.24021v2 Announce Type: replace 
Abstract: Knowledge distillation (KD) is a standard route to compress Large Language Models (LLMs) into compact students, yet most pipelines uniformly apply token-wise loss regardless of teacher confidence. This indiscriminate supervision amplifies noisy, high-entropy signals and is especially harmful under large teacher-student capacity gaps. We introduce SelecTKD, a plug-and-play Selective Token-Weighted distillation framework that shifts the focus from "how to measure divergence" to "where to apply learning". At each step, the student proposes tokens that are verified by the teacher through a robust propose-and-verify procedure with two variants: greedy Top-k and non-greedy Spec-k. Accepted tokens receive full loss, while rejected tokens are masked or down-weighted. This objective-agnostic design works with on- and off-policy data, induces an implicit curriculum quantified by Token Acceptance Rate (TAR), and stabilizes optimization. Across instruction following, mathematical reasoning, code generation, and a VLM setting, SelecTKD consistently improves strong baselines and achieves state-of-the-art results for small models without architectural changes or extra reference models.

SelecTKD: Selective Token-Weighted Knowledge Distillation for LLMs

arXiv:2511.13368v1 Announce Type: new 
Abstract: Large language models (LLMs) perform strongly across tasks and languages, yet how improvements in one task or language affect other tasks and languages and their combinations remains poorly understood. We conduct a controlled PEFT/LoRA study across multiple open-weight LLM families and sizes, treating task and language as transfer axes while conditioning on model family and size; we fine-tune each model on a single task-language source and measure transfer as the percentage-point change versus its baseline score when evaluated on all other task-language target pairs. We decompose transfer into (i) Matched-Task (Cross-Language), (ii) Matched-Language (Cross-Task), and (iii) Cross-Task (Cross-Language) regimes. We uncover two consistent general patterns. First, a pronounced on-task vs. off-task asymmetry: Matched-Task (Cross-Language) transfer is reliably positive, whereas off-task transfer often incurs collateral degradation. Second, a stable donor-recipient structure across languages and tasks (hub donors vs. brittle recipients). We outline implications for risk-aware fine-tuning and model specialisation.

Donors and Recipients: On Asymmetric Transfer Across Tasks and Languages with Parameter-Efficient Fine-Tuning

arXiv:2511.11878v1 Announce Type: new 
Abstract: While large language models (LLMs) show transformative potential in healthcare, their development remains focused on high-resource languages, creating a critical barrier for others as simple translation fails to capture unique clinical and cultural nuances, such as endemic diseases. To address this, we introduce MedPT, the first large-scale, real-world corpus for Brazilian Portuguese, comprising 384,095 authentic question-answer pairs from patient-doctor interactions. The dataset underwent a meticulous multi-stage curation protocol, using a hybrid quantitative-qualitative analysis to filter noise and contextually enrich thousands of ambiguous queries. We further augmented the corpus via LLM-driven annotation, classifying questions into seven semantic types to capture user intent. Our analysis reveals its thematic breadth (3,200 topics) and unique linguistic properties, like the natural asymmetry in patient-doctor communication. To validate its utility, we benchmark a medical specialty routing task: fine-tuning a 1.7B parameter model achieves an outstanding 94\% F1-score on a 20-class setup. Furthermore, our qualitative error analysis shows misclassifications are not random but reflect genuine clinical ambiguities (e.g., between comorbid conditions), proving the dataset's deep semantic richness. We publicly release MedPT to foster the development of more equitable, accurate, and culturally-aware medical technologies for the Portuguese-speaking world.

TEDxTN: A Three-way Speech Translation Corpus for Code-Switched Tunisian Arabic - English

Was this article worth reading? Share it