arXiv:2605.30274v1 Announce Type: new 
Abstract: Document-level translation remains one of the most challenging tasks for large language models, which are constrained by limited context windows that impede global cohesion, while simultaneously suffering from redundant contextual information that degrades translation quality. To address this, we propose a human-like long document translation agent called Loong, which leverages a 3E memory module (Essence-Exemplar-Entity) to store summaries, sentence pairs, and entity records as historical context. Instead of passively attending to all history, Loong performs deep reasoning to adaptively identify the optimal context for translation guidance. Loong optimizes its context policy through reinforcement learning, utilizing preference data derived from its own sampled observe-and-act reasoning trajectories. Empirical evaluations demonstrate that Loong achieves substantial translation quality improvements in English $\Leftrightarrow$ Chinese, German, and French directions, with average gains of up to 13.0 points across the three evaluation metrics. Furthermore, Loong exhibits strong generalization across domains and robustness against contextual noise, while maintaining remarkable stability in ultra-long document translation. Our code is released at https://github.com/YutongWang1216/LoongDocMT.

تم تقديم وكيل جديد لترجمة الوثائق يسمى Loong، والذي يستخدم وحدة ذاكرة 3E لتحسين جودة الترجمة من خلال تخزين السياق التاريخي. يستخدم Loong التفكير العميق لتحديد السياق الأمثل للإرشاد في الترجمة، ويقوم بتحسين سياسته من خلال التعلم المعزز استنادًا إلى بيانات التفضيل المستمدة من مسارات التفكير الخاصة به. تظهر التقييمات التجريبية أن Loong يحقق تحسينات كبيرة في جودة الترجمة بين الإنجليزية والصينية والألمانية والفرنسية.

Se ha presentado un nuevo agente de traducción de documentos llamado Loong, que utiliza un módulo de memoria 3E para mejorar la calidad de la traducción al almacenar el contexto histórico. Loong emplea un razonamiento profundo para seleccionar el contexto óptimo para las traducciones, optimizando su política a través del aprendizaje por refuerzo basado en sus propias trayectorias de razonamiento. Las evaluaciones empíricas indican mejoras significativas en la calidad de la traducción para inglés, chino, alemán y francés.

Un nouvel agent de traduction de documents nommé Loong a été introduit, utilisant un module de mémoire 3E pour améliorer la qualité de la traduction en stockant le contexte historique. Loong utilise un raisonnement approfondi pour sélectionner le contexte optimal pour les traductions, optimisant sa politique grâce à l'apprentissage par renforcement basé sur ses propres trajectoires de raisonnement. Les évaluations empiriques indiquent des améliorations significatives de la qualité de la traduction pour l'anglais, le chinois, l'allemand et le français.

A new document-level translation agent named Loong has been introduced, which utilizes a 3E memory module to enhance translation quality by storing historical context. Loong employs deep reasoning to select optimal context for translations, optimizing its policy through reinforcement learning based on its own reasoning trajectories. Empirical evaluations indicate significant improvements in translation quality for English, Chinese, German, and French.

Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection

arXiv:2606.14122v1 Announce Type: new 
Abstract: Byte-level tokenization enables language models to handle any Unicode input, but models can generate invalid UTF-8 sequences when encountering rare or unseen characters. We investigate the relationship between training scale and UTF-8 generation reliability with a 355M parameter model trained on 80B tokens from a balanced multilingual corpus of English, Japanese, Korean, and Chinese. We introduce multiple evaluation protocols that isolate UTF-8 structural validity from language modeling. UTF-8 validity convergence lags perplexity by a roughly a factor of two: perplexity stabilizes after 2.1B tokens, but UTF-8 validity requires 4.2B tokens. In context-free generation, rare characters achieve higher structural validity than common characters, suggesting over-specialization of frequent character representations. Through experiments, we observed that reliable UTF-8 generation is a distinct capability requiring evaluation beyond perplexity.

دراسة حديثة نُشرت على arXiv تستكشف موثوقية توليد UTF-8 في نماذج اللغة المدركة للبايت، كاشفةً أن الاستقرار في التعقيد يحدث بعد 2.1 مليار رمز، بينما تتطلب صحة UTF-8 4.2 مليار رمز للوصول إلى التقارب. هذه الفجوة تسلط الضوء على التحديات التي تواجهها النماذج عند توليد تسلسلات صالحة لشخصيات نادرة أو غير مرئية.

Un estudio reciente publicado en arXiv investiga la fiabilidad de la generación de UTF-8 en modelos de lenguaje conscientes de bytes, revelando que, mientras que la perplexidad se estabiliza después de 2.1 mil millones de tokens, la validez de UTF-8 requiere 4.2 mil millones de tokens para converger. Esta discrepancia resalta los desafíos que enfrentan los modelos al generar secuencias válidas para caracteres raros o no vistos.

Une étude récente publiée sur arXiv examine la fiabilité de la génération UTF-8 dans les modèles de langage sensibles aux octets, révélant que, tandis que la perplexité se stabilise après 2,1 milliards de jetons, la validité UTF-8 nécessite 4,2 milliards de jetons pour converger. Cette disparité met en lumière les défis auxquels les modèles sont confrontés lors de la génération de séquences valides pour des caractères rares ou non vus.

A recent study published on arXiv investigates the reliability of UTF-8 generation in byte-aware language models, revealing that while perplexity stabilizes after 2.1 billion tokens, UTF-8 validity requires 4.2 billion tokens for convergence. This discrepancy highlights the challenges models face when generating valid sequences for rare or unseen characters.

Beyond Perplexity: UTF-8 Validity in Byte-aware Language Models

arXiv:2606.14278v1 Announce Type: new 
Abstract: Large language models (LLMs) are now widely used as automatic judges for open-ended instruction-following evaluation. This practice is convenient, scalable, and often more semantically aware than reference-based metrics, but it also introduces a new reliability question: does a judge evaluate the quality of an answer, or does it also react to the language in which the comparison is presented? We propose Judge-LS, a lightweight meta-evaluation protocol that transforms LLMBar response-pair items into English, Chinese, and Chinese-English language-switched variants. A reliable judge should preserve its preference under label-preserving language transformations and should not prefer a language when two answers are translation-equivalent. We evaluate four API-accessible judges on the full 419-item LLMBar benchmark, producing 13,408 successful pairwise judgments. Across models, Chinese and language-switched presentations induce 10.7--14.4% preference flips relative to English, and all judges achieve their highest accuracy in English. However, translation-equivalent tie probes do not reveal a systematic English preference: most probes are judged as ties, and non-tie decisions more often favor Chinese. We add confidence intervals, paired significance tests, and an automatic transformation audit with a sensitivity analysis that excludes mechanically flagged high-risk variants. The experiment requires no model training, uses only API calls, and is feasible on modest local hardware.

أظهرت دراسة حديثة أداء نماذج اللغة الكبيرة (LLMs) كقضاة في مهام اتباع التعليمات المفتوحة، حيث تم تقديم بروتوكول يسمى Judge-LS لتقييم ثبات تغيير اللغة. شملت الدراسة 13,408 حكمًا زوجيًا عبر ردود باللغة الإنجليزية والصينية والردود المختلطة، مما كشف عن تفضيل ملحوظ لبعض اللغات في التقييمات.

Un estudio reciente evaluó el rendimiento de los modelos de lenguaje de gran tamaño (LLMs) como jueces en tareas de seguimiento de instrucciones abiertas, introduciendo un protocolo llamado Judge-LS para evaluar la invariancia al cambio de idioma. La investigación involucró 13,408 juicios por pares en respuestas en inglés, chino y variantes de cambio de idioma, revelando una notable preferencia por ciertos idiomas en las evaluaciones.

Une étude récente a évalué la performance des modèles de langage de grande taille (LLMs) en tant que juges dans des tâches d'instruction ouvertes, introduisant un protocole appelé Judge-LS pour évaluer l'invariance au changement de langue. La recherche a impliqué 13 408 jugements par paires dans les réponses en anglais, en chinois et en langues mélangées, révélant une préférence notable pour certaines langues dans les évaluations.

A recent study evaluated the performance of large language models (LLMs) as judges in open-ended instruction-following tasks, introducing a protocol called Judge-LS to assess language-switching invariance. The research involved 13,408 pairwise judgments across English, Chinese, and language-switched responses, revealing a notable preference for certain languages in evaluations.

Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection

Was this article worth reading? Share it

Ready to build your own newsroom?