{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"HPD-Parsing 提出分层并行解码，文档解析吞吐量达 4，752 tokens/s","description":"HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。在公开基准上达到 4，752 tokens/s 的吞吐量，比自回归基线提升 3.06 倍，同时保持有竞争力的解析精度。该方法为高效统一文档解析开辟了新方向。","url":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/","mainEntityOfPage":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/","datePublished":"2026-07-21T00:00:00.000Z","dateModified":"2026-07-21T00:00:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://arxiv.org/abs/2607.18839","https://aihot.virxact.com/items/cmrvls9qg03jvbihb0wv6okzw"],"canonicalUrl":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/","dateCreated":"2026-07-21T00:00:00.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"arXiv source article","url":"https://arxiv.org/abs/2607.18839","datePublished":"2026-07-21T00:00:00.000Z","provider":{"@type":"Organization","name":"arXiv","url":"https://arxiv.org/abs/2607.18839"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrvls9qg03jvbihb0wv6okzw","datePublished":"2026-07-21T00:00:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrvls9qg03jvbihb0wv6okzw"}}],"aggregationSource":"HuggingFace Daily Papers（社区热门论文）","originalPublisher":{"name":"arXiv","url":"https://arxiv.org/abs/2607.18839"},"article":{"id":"cmrvls9qg03jvbihb0wv6okzw","slug":"cmrvls9qg03jvbihb0wv6okzw","url":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/","title":"HPD-Parsing 提出分层并行解码，文档解析吞吐量达 4，752 tokens/s","title_en":"HPD-Parsing： Hierarchical Parallel Document Parsing","summary":"HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。在公开基准上达到 4，752 tokens/s 的吞吐量，比自回归基线提升 3.06 倍，同时保持有竞争力的解析精度。该方法为高效统一文档解析开辟了新方向。","source":"HuggingFace Daily Papers（社区热门论文）","sourceUrl":"https://arxiv.org/abs/2607.18839","aiHotUrl":"https://aihot.virxact.com/items/cmrvls9qg03jvbihb0wv6okzw","publishedAt":"2026-07-21T00:00:00.000Z","category":"论文研究","score":54,"selected":false,"articleBody":["arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.","Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs ：https://info.arxiv.org/labs/index.html."],"articleImages":[{"sourceUrl":"https://arxiv.org/static/base/1.0.1/images/funders/simons-foundation.png","alt":"Simons Foundation","afterParagraph":1,"url":"/media/articles/cmrvls9qg03jvbihb0wv6okzw/e2d7f38d62f5ca91.png"},{"sourceUrl":"https://arxiv.org/static/base/1.0.1/images/funders/simons-foundation-international.png","alt":"Simons Foundation International","afterParagraph":1,"url":"/media/articles/cmrvls9qg03jvbihb0wv6okzw/1d56e29c5557cbdc.png"}],"mediaStatus":"ok","articleBodyZh":["arXivLabs 是一个框架，允许合作者直接在我们的网站上开发和分享新的 arXiv 功能。","有一个可以为 arXiv 社区增加价值的项目想法吗？了解更多关于 arXivLabs 的信息：https://info.arxiv.org/labs/index.html。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-07-23T06:49:19.092Z","sourceHash":"23e2c2927703f8ed","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["论文研究","HuggingFace Daily Papers（社区热门论文）"],"translations":{"zh-CN":{"title":"HPD-Parsing 提出分层并行解码，文档解析吞吐量达 4，752 tokens/s","summary":"HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。在公开基准上达到 4，752 tokens/s 的吞吐量，比自回归基线提升 3.06 倍，同时保持有竞争力的解析精度。该方法为高效统一文档解析开辟了新方向。","category":"论文研究","source":"arXiv","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing 提出分层并行解码，文档解析吞吐量达 4，752 tokens/s - Aioga AI资讯","description":"HPD-Parsing 用分层并行解码替代全页自回归生成：主布局分支协调全局结构并动态分配块级内容解码到并发分支，渐进式多 token 预测（P-MTP）进一步减少各分支解码步数。在公开基准上达到 4，752 tokens/s 的吞吐量，比自回归基线提升 3.06 倍，同时保持有竞争力的解析精度。该方法为高效统一文档解析开辟了新方向。","url":"https://www.aioga.com/news/cmrvls9qg03jvbihb0wv6okzw/"},"en":{"title":"HPD-Parsing proposes hierarchical parallel decoding, achieving a document parsing throughput of 4,752 tokens/s","summary":"HPD-Parsing replaces full-page autoregressive generation with hierarchical parallel decoding: the main layout branch coordinates the global structure and dynamically allocates block-level content decoding to concurrent branches, while progressive multi-token prediction (P-MTP) further reduces the decoding steps of each branch. On public benchmarks, it achieves a throughput of 4,752 tokens/s, 3.06 times higher than the autoregressive baseline, while maintaining competitive parsing accuracy. This approach opens a new direction for efficient unified document parsing.","category":"Research","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing proposes hierarchical parallel decoding, achieving a document parsing throughput of 4,752 tokens/s - Aioga AI News","description":"HPD-Parsing replaces full-page autoregressive generation with hierarchical parallel decoding: the main layout branch coordinates the global structure and dynamically allocates bloc...","url":"https://www.aioga.com/en/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:42:07.426Z"},"ja":{"title":"HPD-Parsing は階層並列デコーディングを提出し、ドキュメント解析のスループットは 4,752 tokens/s に達する","summary":"HPD-Parsing における階層的並列デコーディングによる全ページ自己回帰生成の代替：主レイアウトブランチが全体構造を調整し、ブロックレベルのコンテンツデコーディングを並行ブランチに動的に割り当て、漸進的マルチトークン予測（P-MTP）が各ブランチのデコーディングステップ数をさらに削減します。公開ベンチマークでは 4,752 tokens/s のスループットを達成し、自己回帰ベースラインに比べて 3.06 倍向上しつつ、競争力のある解析精度を維持しています。この方法は、効率的な統一型ドキュメント解析に新たな方向性を開きます。","category":"論文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing は階層並列デコーディングを提出し、ドキュメント解析のスループットは 4,752 tokens/s に達する - Aioga AIニュース","description":"HPD-Parsing における階層的並列デコーディングによる全ページ自己回帰生成の代替：主レイアウトブランチが全体構造を調整し、ブロックレベルのコンテンツデコーディングを並行ブランチに動的に割り当て、漸進的マルチトークン予測（P-MTP）が各ブランチのデコーディングステップ数をさらに削減します。公開ベンチマークでは 4,752 tokens/s のスループ...","url":"https://www.aioga.com/ja/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:42:18.705Z"},"ko":{"title":"HPD-Parsing은 계층적 병렬 디코딩을 제안하며, 문서 파싱 처리량은 4,752 토큰/초에 달합니다","summary":"HPD-Parsing은 전체 페이지 자회귀 생성을 대신하여 계층 병렬 디코딩을 사용합니다: 주 레이아웃 분기가 전체 구조를 조정하고 블록 단위 콘텐츠 디코딩을 동시 분기로 동적으로 할당하며, 점진적 다중 토큰 예측(P-MTP)은 각 분기 디코딩 단계를 더욱 줄입니다. 공개 벤치마크에서 초당 4,752 토큰의 처리량을 달성하여 자회귀 기준보다 3.06배 향상되었으며, 경쟁력 있는 파싱 정밀도를 유지합니다. 이 방법은 효율적인 통합 문서 파싱에 새로운 방향을 열었습니다.","category":"연구","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing은 계층적 병렬 디코딩을 제안하며, 문서 파싱 처리량은 4,752 토큰/초에 달합니다 - Aioga AI 뉴스","description":"HPD-Parsing은 전체 페이지 자회귀 생성을 대신하여 계층 병렬 디코딩을 사용합니다: 주 레이아웃 분기가 전체 구조를 조정하고 블록 단위 콘텐츠 디코딩을 동시 분기로 동적으로 할당하며, 점진적 다중 토큰 예측(P-MTP)은 각 분기 디코딩 단계를 더욱 줄입니다. 공개 벤치마크에서 초당 4,752 토큰의 처리량을 달성...","url":"https://www.aioga.com/ko/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:43:10.171Z"},"es":{"title":"HPD-Parsing propone decodificación paralela por niveles, alcanzando un rendimiento de análisis de documentos de 4.752 tokens/s","summary":"HPD-Parsing utiliza decodificación paralela en capas en lugar de generación autoregresiva completa de la página: la rama de diseño principal coordina la estructura global y asigna dinámicamente la decodificación de contenido a nivel de bloque a ramas concurrentes, la predicción progresiva de múltiples tokens (P-MTP) reduce aún más el número de pasos de decodificación en cada rama. En benchmarks públicos, alcanza un rendimiento de 4,752 tokens/s, aumentando 3.06 veces en comparación con la línea base autoregresiva, al mismo tiempo que mantiene una precisión de análisis competitiva. Este enfoque abre una nueva dirección para el análisis de documentos unificado y eficiente.","category":"Investigación","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing propone decodificación paralela por niveles, alcanzando un rendimiento de análisis de documentos de 4.752 tokens/s - Aioga Noticias de IA","description":"HPD-Parsing utiliza decodificación paralela en capas en lugar de generación autoregresiva completa de la página: la rama de diseño principal coordina la estructura global y asigna...","url":"https://www.aioga.com/es/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:42:58.526Z"},"fr":{"title":"HPD-Parsing propose un décodage parallèle hiérarchique, le débit d'analyse des documents atteint 4 752 tokens/s","summary":"HPD-Parsing utilise le décodage parallèle en couches pour remplacer la génération autoregressive sur toute la page : la branche de mise en page principale coordonne la structure globale et répartit dynamiquement le décodage du contenu au niveau des blocs vers des branches concurrentes, tandis que la prédiction progressive de multiples tokens (P-MTP) réduit encore le nombre d'étapes de décodage pour chaque branche. Sur des benchmarks publics, le débit atteint 4 752 tokens/s, soit une amélioration de 3,06 fois par rapport à la base autoregressive, tout en maintenant une précision de parsing compétitive. Cette méthode ouvre une nouvelle voie pour l'analyse efficace et unifiée des documents.","category":"Recherche","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing propose un décodage parallèle hiérarchique, le débit d'analyse des documents atteint 4 752 tokens/s - Aioga Actualités IA","description":"HPD-Parsing utilise le décodage parallèle en couches pour remplacer la génération autoregressive sur toute la page : la branche de mise en page principale coordonne la structure gl...","url":"https://www.aioga.com/fr/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:43:56.557Z"},"de":{"title":"HPD-Parsing schlägt eine hierarchische parallele Decodierung vor, die Dokumentenparsing-Durchsatz beträgt 4.752 Token/s","summary":"HPD-Parsing ersetzt die autoregressive Generierung der gesamten Seite durch hierarchisches paralleles Decodieren: Der Hauptlayoutzweig koordiniert die globale Struktur und verteilt dynamisch die Blockinhalte zum Decodieren auf parallele Zweige. Progressives Multi-Token-Vorhersagen (P-MTP) reduziert weiter die Decodierschritte der einzelnen Zweige. Auf öffentlichen Benchmarks wird eine Durchsatzrate von 4.752 Tokens/s erreicht, was im Vergleich zur autoregressiven Basislinie eine Steigerung um das 3,06-fache darstellt, während gleichzeitig eine wettbewerbsfähige Parsing-Genauigkeit beibehalten wird. Diese Methode eröffnet einen neuen Weg für effizientes, einheitliches Dokumentenparsing.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing schlägt eine hierarchische parallele Decodierung vor, die Dokumentenparsing-Durchsatz beträgt 4.752 Token/s - Aioga KI-News","description":"HPD-Parsing ersetzt die autoregressive Generierung der gesamten Seite durch hierarchisches paralleles Decodieren: Der Hauptlayoutzweig koordiniert die globale Struktur und verteilt...","url":"https://www.aioga.com/de/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:43:53.922Z"},"pt-BR":{"title":"HPD-Parsing propõe decodificação paralela em camadas, com rendimento de análise de documentos de 4.752 tokens/s","summary":"HPD-Parsing substitui a geração autoregressiva de página inteira por decodificação paralela em camadas: o ramo de layout principal coordena a estrutura global e distribui dinamicamente a decodificação de conteúdo em nível de bloco para ramos concorrentes, a previsão progressiva de múltiplos tokens (P-MTP) reduz ainda mais o número de etapas de decodificação em cada ramo. Em benchmarks públicos, atinge uma taxa de transferência de 4.752 tokens/s, 3,06 vezes maior que a linha de base autoregressiva, mantendo ao mesmo tempo precisão de parsing competitiva. Este método abre um novo caminho para parsing unificado de documentos de forma eficiente.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing propõe decodificação paralela em camadas, com rendimento de análise de documentos de 4.752 tokens/s - Aioga Notícias de IA","description":"HPD-Parsing substitui a geração autoregressiva de página inteira por decodificação paralela em camadas: o ramo de layout principal coordena a estrutura global e distribui dinamicam...","url":"https://www.aioga.com/pt-BR/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:44:48.235Z"},"ru":{"title":"HPD-Parsing предлагает иерархическую параллельную декодировку, пропускная способность анализа документов достигает 4 752 токенов/с","summary":"HPD-Parsing использует иерархическую параллельную декодировку вместо полной автогенерации страниц: основной макетный блок координирует глобальную структуру и динамически распределяет декодирование содержимого на блочном уровне в конкурентные ветви, а поэтапное прогнозирование нескольких токенов (P-MTP) дополнительно уменьшает количество шагов декодирования для каждой ветви. На общедоступных бенчмарках достигается пропускная способность 4 752 токенов/с, что в 3,06 раза выше базовой автогенерации, при этом сохраняя конкурентоспособную точность анализа. Этот метод открывает новое направление для эффективного унифицированного анализа документов.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing предлагает иерархическую параллельную декодировку, пропускная способность анализа документов достигает 4 752 токенов/с - Aioga Новости ИИ","description":"HPD-Parsing использует иерархическую параллельную декодировку вместо полной автогенерации страниц: основной макетный блок координирует глобальную структуру и динамически распределя...","url":"https://www.aioga.com/ru/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:44:40.677Z"},"ar":{"title":"HPD-Parsing يقترح فك الترميز المتوازي الطبقي، ويصل معدل معالجة المستندات إلى 4,752 رمز في الثانية","summary":"HPD-Parsing يستخدم فك التشفير المتوازي الهرمي بدلاً من التوليد الذاتي التكراري للصفحة بالكامل: يقوم الفرع الرئيسي لتخطيط الصفحة بتنسيق الهيكل العام وتوزيع فك تشفير المحتوى على مستوى الكتل ديناميكيًا إلى الفروع المتزامنة، كما يقلل التنبؤ التدريجي بعدة رموز (P-MTP) من عدد خطوات فك التشفير لكل فرع. على المعايير العامة، يصل معدل الإنتاجية إلى 4،752 رمزًا/ثانية، أي أعلى بمقدار 3.06 مرة من الأساس التكراري الذاتي، مع الحفاظ على دقة التحليل التنافسية. تفتح هذه الطريقة اتجاهًا جديدًا لتحليل المستندات الموحد بشكل فعال.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing يقترح فك الترميز المتوازي الطبقي، ويصل معدل معالجة المستندات إلى 4,752 رمز في الثانية - Aioga أخبار الذكاء الاصطناعي","description":"HPD-Parsing يستخدم فك التشفير المتوازي الهرمي بدلاً من التوليد الذاتي التكراري للصفحة بالكامل: يقوم الفرع الرئيسي لتخطيط الصفحة بتنسيق الهيكل العام وتوزيع فك تشفير المحتوى على مستو...","url":"https://www.aioga.com/ar/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:45:41.294Z"},"hi":{"title":"HPD-Parsing ने स्तरीकृत समानांतर डिकोडिंग का प्रस्ताव किया, दस्तावेज़ पार्सिंग थ्रूपुट 4,752 टोकन/सेकंड तक पहुँच गया","summary":"HPD-Parsing पूरे पेज आत्म-प्रतिगामी उत्पादन को प्रतिस्थापित करने के लिए परत-दर-परत समांतर डिकोडिंग का उपयोग करता है: मुख्य लेआउट शाखा वैश्विक संरचना का समन्वय करती है और ब्लॉक-स्तरीय सामग्री डिकोडिंग को समवर्ती शाखाओं में गतिशील रूप से वितरित करती है, प्रगतिशील मल्टी-टोकन प्रेडिक्शन (P-MTP) प्रत्येक शाखा के डिकोडिंग स्टेप्स को और कम करता है। सार्वजनिक बेंचमार्क पर यह 4,752 tokens/s की थ्रूपुट तक पहुँचता है, जो आत्म-प्रतिगामी बेसलाइन की तुलना में 3.06 गुना अधिक है, साथ ही प्रतिस्पर्धी विश्लेषण सटीकता बनाए रखता है। यह तरीका कुशल एकीकृत दस्तावेज़ पार्सिंग के लिए नया मार्ग खोलता है।","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing ने स्तरीकृत समानांतर डिकोडिंग का प्रस्ताव किया, दस्तावेज़ पार्सिंग थ्रूपुट 4,752 टोकन/सेकंड तक पहुँच गया - Aioga AI समाचार","description":"HPD-Parsing पूरे पेज आत्म-प्रतिगामी उत्पादन को प्रतिस्थापित करने के लिए परत-दर-परत समांतर डिकोडिंग का उपयोग करता है: मुख्य लेआउट शाखा वैश्विक संरचना का समन्वय करती है और ब्लॉक-स्तर...","url":"https://www.aioga.com/hi/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:45:40.402Z"},"it":{"title":"HPD-Parsing propone la decodifica parallela gerarchica, con un throughput di analisi dei documenti di 4.752 token/s","summary":"HPD-Parsing sostituisce la generazione autoregressiva a pagina intera con la decodifica parallela gerarchica: il ramo principale del layout coordina la struttura globale e assegna dinamicamente la decodifica del contenuto a livello di blocco ai rami concorrenti, mentre la previsione progressiva multi-token (P-MTP) riduce ulteriormente il numero di passaggi di decodifica per ciascun ramo. Sui benchmark pubblici, raggiunge una velocità di 4.752 token/s, migliorando di 3,06 volte rispetto al modello autoregressivo di base, mantenendo al contempo una precisione di parsing competitiva. Questo metodo apre una nuova direzione per un parsing efficiente e unificato dei documenti.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing propone la decodifica parallela gerarchica, con un throughput di analisi dei documenti di 4.752 token/s - Aioga Notizie IA","description":"HPD-Parsing sostituisce la generazione autoregressiva a pagina intera con la decodifica parallela gerarchica: il ramo principale del layout coordina la struttura globale e assegna...","url":"https://www.aioga.com/it/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:46:27.441Z"},"nl":{"title":"HPD-Parsing stelt hiërarchische parallelle decodering voor, documentparserdoorvoer bereikt 4.752 tokens/s","summary":"HPD-Parsing vervangt volledige pagina autoregressieve generatie door gelaagde parallelle decodering: de hoofdlay-outtak coördineert de globale structuur en wijst dynamisch blokniveau-inhoudsdecodering toe aan gelijktijdige takken, en stapsgewijze multi-token voorspelling (P-MTP) vermindert het aantal decoderingstappen per tak verder. Op openbare benchmarks wordt een doorvoer van 4.752 tokens/s bereikt, 3,06 keer hoger dan de autoregressieve baseline, terwijl een concurrerende parsingsnauwkeurigheid behouden blijft. Deze methode opent een nieuwe richting voor efficiënte uniforme documentparsing.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing stelt hiërarchische parallelle decodering voor, documentparserdoorvoer bereikt 4.752 tokens/s - Aioga AI-nieuws","description":"HPD-Parsing vervangt volledige pagina autoregressieve generatie door gelaagde parallelle decodering: de hoofdlay-outtak coördineert de globale structuur en wijst dynamisch bloknive...","url":"https://www.aioga.com/nl/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:46:27.758Z"},"tr":{"title":"HPD-Parsing, katmanlı paralel kod çözmeyi önerir ve belge çözümleme verimliliği 4.752 token/s'ye ulaşır","summary":"HPD-Parsing, tüm sayfa oto-regresif üretimi yerine katmanlı paralel çözümlemeyi kullanır: Ana düzen dalı global yapıyı koordine eder ve blok düzeyindeki içerik çözümlemesini eşzamanlı dallara dinamik olarak dağıtır, kademeli çok token tahmini (P-MTP) ise her dalın çözümleme adımlarını daha da azaltır. Açık benchmark üzerinde saniyede 4.752 token verimlilik elde ederek oto-regresif tabana göre 3,06 kat artış sağlarken, rekabetçi çözümleme doğruluğunu korur. Bu yöntem, verimli ve birleşik bir belge çözümlemesi için yeni bir yön açar.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing, katmanlı paralel kod çözmeyi önerir ve belge çözümleme verimliliği 4.752 token/s'ye ulaşır - Aioga AI Haberleri","description":"HPD-Parsing, tüm sayfa oto-regresif üretimi yerine katmanlı paralel çözümlemeyi kullanır: Ana düzen dalı global yapıyı koordine eder ve blok düzeyindeki içerik çözümlemesini eşzama...","url":"https://www.aioga.com/tr/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:47:10.194Z"},"vi":{"title":"HPD-Parsing đề xuất giải mã song song theo tầng, thông lượng phân tích tài liệu đạt 4.752 token/s","summary":"HPD-Parsing sử dụng giải mã song song theo phân cấp thay cho sinh tự hồi quy toàn trang: nhánh bố cục chính phối hợp cấu trúc toàn cục và phân bổ động việc giải mã nội dung cấp khối cho các nhánh đồng thời, dự đoán nhiều token tiến tiến (P-MTP) tiếp tục giảm số bước giải mã của các nhánh. Trên các chuẩn công khai đạt được thông lượng 4.752 tokens/s, cao hơn 3,06 lần so với baseline tự hồi quy, đồng thời duy trì độ chính xác phân tích cạnh tranh. Phương pháp này mở hướng mới cho phân tích tài liệu thống nhất hiệu quả.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing đề xuất giải mã song song theo tầng, thông lượng phân tích tài liệu đạt 4.752 token/s - Tin tức AI Aioga","description":"HPD-Parsing sử dụng giải mã song song theo phân cấp thay cho sinh tự hồi quy toàn trang: nhánh bố cục chính phối hợp cấu trúc toàn cục và phân bổ động việc giải mã nội dung cấp khố...","url":"https://www.aioga.com/vi/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:47:14.560Z"},"id":{"title":"HPD-Parsing mengusulkan dekoding paralel bertingkat, throughput pemrosesan dokumen mencapai 4.752 token/s","summary":"HPD-Parsing menggunakan decoding paralel bertingkat sebagai pengganti generasi autoregresif seluruh halaman: cabang tata letak utama mengoordinasikan struktur global dan secara dinamis mendistribusikan decoding konten tingkat blok ke cabang konkuren, prediksi multi-token secara progresif (P-MTP) lebih lanjut mengurangi jumlah langkah decoding di setiap cabang. Pada benchmark publik mencapai throughput 4.752 token/s, meningkat 3,06 kali dibandingkan baseline autoregresif, sambil tetap mempertahankan akurasi parsing yang kompetitif. Metode ini membuka arah baru untuk parsing dokumen terpadu yang efisien.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing mengusulkan dekoding paralel bertingkat, throughput pemrosesan dokumen mencapai 4.752 token/s - Berita AI Aioga","description":"HPD-Parsing menggunakan decoding paralel bertingkat sebagai pengganti generasi autoregresif seluruh halaman: cabang tata letak utama mengoordinasikan struktur global dan secara din...","url":"https://www.aioga.com/id/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:47:58.021Z"},"th":{"title":"HPD-Parsing นำเสนอการถอดรหัสแบบขนานเป็นชั้น ความเร็วในการประมวลผลเอกสารถึง 4,752 token/s","summary":"HPD-Parsing ใช้การถอดรหัสแบบลำดับชั้นขนานแทนการสร้างแบบอัตนัยทั้งหน้า: สาขาเลย์เอาต์หลักประสานโครงสร้างทั่วทั้งหน้าและจัดสรรการถอดรหัสเนื้อหาระดับบล็อกไปยังสาขาขนานอย่างมีพลวัต การทำนายหลายโทเค็นแบบก้าวหน้า (P-MTP) ลดจำนวนขั้นตอนการถอดรหัสของแต่ละสาขาได้เพิ่มเติม ที่มาตรฐานแบบเปิดมีอัตราการประมวลผล 4,752 โทเค็น/วินาที เพิ่มขึ้น 3.06 เท่าจากฐานอัตนัย ในขณะที่ยังคงรักษาความแม่นยำในการวิเคราะห์ที่แข่งขันได้ วิธีนี้ได้เปิดแนวทางใหม่สำหรับการวิเคราะห์เอกสารแบบรวมที่มีประสิทธิภาพ","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing นำเสนอการถอดรหัสแบบขนานเป็นชั้น ความเร็วในการประมวลผลเอกสารถึง 4,752 token/s - ข่าว AI Aioga","description":"HPD-Parsing ใช้การถอดรหัสแบบลำดับชั้นขนานแทนการสร้างแบบอัตนัยทั้งหน้า: สาขาเลย์เอาต์หลักประสานโครงสร้างทั่วทั้งหน้าและจัดสรรการถอดรหัสเนื้อหาระดับบล็อกไปยังสาขาขนานอย่างมีพลวัต การ...","url":"https://www.aioga.com/th/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:48:09.418Z"},"pl":{"title":"HPD-Parsing wprowadza hierarchiczne równoległe dekodowanie, przepustowość analizy dokumentów osiąga 4 752 tokenów/s","summary":"HPD-Parsing zastępuje generowanie autoregresyjne całej strony hierarchicznym równoległym dekodowaniem: główna gałąź układu koordynuje strukturę globalną i dynamicznie przydziela dekodowanie zawartości na poziomie bloków do gałęzi równoległych, a progresywne przewidywanie wielu tokenów (P-MTP) dodatkowo zmniejsza liczbę kroków dekodowania każdej gałęzi. Na publicznych benchmarkach osiąga przepustowość 4 752 tokenów/s, co stanowi wzrost o 3,06 razy w porównaniu do bazowego modelu autoregresyjnego, przy zachowaniu konkurencyjnej dokładności parsowania. Ta metoda otwiera nowy kierunek w efektywnym, zunifikowanym parsowaniu dokumentów.","category":"论文研究","source":"HuggingFace Daily Papers（社区热门论文）","aggregationSource":"HuggingFace Daily Papers（社区热门论文）","pageTitle":"HPD-Parsing wprowadza hierarchiczne równoległe dekodowanie, przepustowość analizy dokumentów osiąga 4 752 tokenów/s - Aioga Wiadomości AI","description":"HPD-Parsing zastępuje generowanie autoregresyjne całej strony hierarchicznym równoległym dekodowaniem: główna gałąź układu koordynuje strukturę globalną i dynamicznie przydziela de...","url":"https://www.aioga.com/pl/news/cmrvls9qg03jvbihb0wv6okzw/","contentTranslated":true,"sourceHash":"c60e8cd62f4748a2","translatedAt":"2026-07-23T02:48:56.915Z"}}}}