{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-08-11T09:21:12.743Z","headline":"Meta 广告排序的多阶段序列模型：从用户序列到 LLM 式扩展定律","description":"Meta 推出多阶段序列模型，将离线用户建模与在线排序解耦，并采用密集 token 化与目标感知注意力，使序列学习具备可预测的 LLM 式扩展定律。该架构已为 Instagram 转化率带来 6% 的累计提升，Facebook 转化率提升 3%、广告点击量提升 3.5%，并成为 Meta 生成式广告推荐模型（GEM）的核心组件。","url":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","mainEntityOfPage":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","datePublished":"2026-08-05T19:20:20.000Z","dateModified":"2026-08-05T19:20:20.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking","https://aihot.virxact.com/items/cmsgi5rk106bero5qx58elyce"],"canonicalUrl":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","directAnswer":{"@type":"Answer","text":"Meta 介绍多阶段序列模型，将高计算量的离线用户建模与轻量在线排序分离，并结合密集 token 化和目标感知注意力，以可预测的扩展规律提升广告排序。","url":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","dateCreated":"2026-08-05T19:20:20.000Z","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":"engineering.fb.com source article","url":"https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking","datePublished":"2026-08-05T19:20:20.000Z","provider":{"@type":"Organization","name":"engineering.fb.com","url":"https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmsgi5rk106bero5qx58elyce","datePublished":"2026-08-05T19:20:20.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmsgi5rk106bero5qx58elyce"}}],"aggregationSource":"Meta Engineering Blog（RSS）","originalPublisher":{"name":"engineering.fb.com","url":"https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking"},"geoDeepAnswer":null,"article":{"id":"cmsgi5rk106bero5qx58elyce","slug":"cmsgi5rk106bero5qx58elyce","url":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","title":"Meta 广告排序的多阶段序列模型：从用户序列到 LLM 式扩展定律","title_en":"From User Sequences to Scaling Laws： A Multi-Stage Architecture for Meta's Ads Ranking","summary":"Meta 推出多阶段序列模型，将离线用户建模与在线排序解耦，并采用密集 token 化与目标感知注意力，使序列学习具备可预测的 LLM 式扩展定律。该架构已为 Instagram 转化率带来 6% 的累计提升，Facebook 转化率提升 3%、广告点击量提升 3.5%，并成为 Meta 生成式广告推荐模型（GEM）的核心组件。","source":"Meta Engineering Blog（RSS）","sourceUrl":"https://engineering.fb.com/2026/08/05/ml-applications/from-user-sequences-to-scaling-laws-a-multi-stage-architecture-for-metas-ads-ranking","aiHotUrl":"https://aihot.virxact.com/items/cmsgi5rk106bero5qx58elyce","publishedAt":"2026-08-05T19:20:20.000Z","category":"产品更新","score":58,"selected":false,"articleBody":["Every day, Meta’s recommendation platforms handle billions of user interactions, generating rich temporal signals that capture individual preferences and intent across products, ads, and content. In our 2024 post on sequence learning for ads recommendations ：https://engineering.fb.com/2024/11/19/data-infrastructure/sequence-learning-personalized-ads-recommendations/ , we showed how modeling the order and timing of user actions ( rather than relying on static, manually engineered sparse features) produces richer, sequence-aware representations of user interests and ad preferences.","This post goes a step further, introducing two architectural breakthroughs that let us scale sequence learning advancements from foundational innovations into a production platform with predictable, LLM-style scaling laws: (1) a multi-stage sequence model that decouples heavy offline user modeling from lightweight online ranking tasks and (2) a learning technique based on dense tokenization and target-aware attention that efficiently learns feature interactions directly from data.","Together with our broader model innovations, these advancements have contributed to a cumulative lift of 6% in conversions on Instagram, 3% in conversions on Facebook and 3.5% in ad clicks on Facebook. This unified platform for sequence modeling is a core component of Meta’s Generative Ads Recommendation Model (GEM) ：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/ , helps to harness the comprehensive user behavioral understanding of this learning paradigm to maximize the benefit to advertisers.","Ads recommendation systems must retrieve and rank thousands of ads within milliseconds, processing millions of candidates per second. To manage this scale, some approaches to sequence models rely on hybrid model configurations where a specific model processes user event sequences and another model handles sparse feature interactions.","While effective at meeting production demands, this hybrid approach has potential tradeoffs:","Scaling both temporal sequence lengths and the transformer models that process them can turn the tradeoffs of the hybrid approach into a bottleneck, limiting the ability to improve the ads experience of users and the performance of advertisers’ campaigns.","We’ve made two fundamental architectural breakthroughs in sequence learning that resolve the core tension between model complexity and serving efficiency: (1) a multi-stage sequence model that decouples offline user modeling from online ranking and (2) a dense tokenization with target-aware attention learning paradigm. Together, they provide a flexible production strategy that helps generalize sequence learning models and establish an LLM-style scaling law that predictably balances model performance with compute.","To address scaling efficiency, a multi-stage model has been developed that enables scaling of a transformer-based sequence model in a compute efficient manner. Separating the sequence model into two complementary stages (upstream/offline user modeling and downstream/online ranking), enables model capacity to scale so that performance keeps improving without proportional increases in serving resources.","In Figure 1 , the left panel shows the offline user model. It processes long user histories asynchronously and produces cached embeddings that capture deep behavioral patterns. The right panel shows the online ranking model that combines these cached representations with real time ad candidate signals to produce the final ranking. The arrow between the two stages carries the user feature embeddings from offline → online ranking models.","User-side features are processed asynchronously using deep transformer upstream models. These models scale to several transformer layers with sequence lengths in the thousands and generate embeddings that are precomputed and cached at the user level. The upstream model strictly separates user features from ad and context features to ensure user embeddings remain independent of any particular ad candidate.","The offline user model representations are complemented with online ranking models that use fresh user signals and ad candidate information for real time ranking. This stage is optimized for speed, meeting strict latency budgets while leveraging the deep representations computed offline.","Separating the sequence modeling system into two distinct, yet complementary, stages enables an increase in model complexity along a scaling curve for the Offline User Model without causing a spike in serving costs for the Online Ranking Models .","This tokenization approach integrates sparse features with sequential behavioral data into a single dense vocabulary, enabling attention mechanisms to discover interactions independently. Unlike traditional recommendation systems, which rely on manually engineered representations to capture sparse cross-feature interactions, this approach lets the model learn those interactions directly from the data.","Tokenized sparse features and ad candidate information are fused with user behavior sequences, then processed by a memory-efficient form of multi-head attention that lets each layer weigh a user’s past behaviors against the specific ad being scored. Stacking multiple aligned attention blocks with stable attention distributions allows each layer to capture higher-order interactions between the target ad and the user’s historical behavior, progressively distilling long sequences into compact representations.","When running on real-world ads traffic, the multi-stage sequence model demonstrates the emergence of predictable scaling laws for ads recommendations that are analogous to those observed in large language models. Performance improvements follow a log-linear relationship with respect to compute, with a marked improvement in scaling efficiency over other transformer-based sequence models. Figure 2 conveys these scaling properties by showing the relationship between compute (FLOPs) and model performance (measured by normalized entropy, NE) across several dimensions: model depth, content/semantic enrichment, model width, and sequence length.","Unlike LLMs, which process dense and continuous text, ads recommendation systems must integrate sparse ID features with temporal user sequences. The fact that LLM-style scaling emerged despite the structural differences provides a strong indicator of model architectural fit for further sequence learning applications.","We have identified four levers that we anticipate will help unlock the frontier of the scaling law:","Optimal performance requires balanced growth across model depth, width and sequence lengths. If scaling only occurs on a single axis, the other axes will likely bottleneck the performance improvements, potentially leading to diminishing returns. This mirrors findings from LLM scaling law research, a principle we call the scaling synergy principle .","The multi-stage architecture provides a tunable lever to scale either the offline or online model up/down. Scaling the online ranking model drives steeper improvements per unit of compute that is bounded by serving/request time requirements. Scaling the offline model (shown in Figure 2) follows a more gradual curve, but its async inference avoids latency constraints, allowing scale in at an unhindered rate.","Performance continues to improve as sequences get longer, but an impactful finding is that sequence diversity beats sequence homogeneity. A balanced mix of action types (e.g., views, clicks, conversions) yields better results than sequences composed of a single action type. This finding suggests that a diverse mix of engagement types and broad temporal coverage produce richer behavioral representations of users than homogeneous sequences of high signal actions in isolation.","Semantic content features from foundation models complement traditional collaborative filtering (i.e. which users interacted with which items) signals. They are especially helpful in cold-start scenarios (e.g., new ads or advertisers with limited historical engagement data). By addressing this persistent challenge of recommendation systems, we improve overall signal coverage to a fundamental sparse problem in recommendation systems.","The multi-stage sequence modeling architecture is delivering impact across three dimensions:","By modeling thousands of user event sequences (e.g., clicks, views, and purchases) the offline model generates highly nuanced user representations. This depth of behavioral understanding improves ad relevance and conversion rates across Meta’s Family of Apps. Together with our broader modeling innovations, these sequence-derived representations drove a cumulative lift of 6% in conversions on Instagram, 3% on Facebook and 3.5% in ad clicks on Facebook.","The two-stage design delivers performance improvements with greater compute efficiency compared to hybrid approaches. Initial evaluations improved ads ranking quality with minimal impact to serving resources, confirming that model complexity and production efficiency can scale together.","As a core part of GEM ：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/ , this model architecture for sequence learning has been designed for generalization, where the same multi-stage backbone and scaling properties can extend to any ads ranking task with minimal adaptation and overhead.","The sequence model scaling law shows no signs of saturation. With architectural parity achieved, scaling model complexity can draw on techniques proven in the LLM domain (e.g., mixture-of-experts, cross-user compute sharing, advanced attention mechanisms) with potential to continually scale at the optimal performance/efficiency tradeoff.","A detailed technical publication of this model architecture and its scaling properties is available in our paper, “ LLaTTE: Scaling Laws for Multi-Stage Sequence Modeling in Large-Scale Ads Recommendation ：https://arxiv.org/pdf/2601.20083 .”","Meta believes in building community through open source technology. Explore our latest projects in Artificial Intelligence, Data Infrastructure, Development Tools, Front End, Languages, Platforms, Security, Virtual Reality, and more.","Engineering at Meta is a technical news resource for engineers interested in how we solve large-scale technical challenges at Meta."],"articleImages":[{"sourceUrl":"https://engineering.fb.com/wp-content/uploads/2026/07/image3.jpg?w=580&h=326&crop=1","alt":"","afterParagraph":26,"url":"/media/articles/cmsgi5rk106bero5qx58elyce/8805d4351ccae466.webp"},{"sourceUrl":"https://engineering.fb.com/wp-content/uploads/2026/06/10-Years-Meta-Python-HERO-large.png?w=580&h=326&crop=1","alt":"","afterParagraph":26,"url":"/media/articles/cmsgi5rk106bero5qx58elyce/85fcdba8ca3e4cf2.webp"},{"sourceUrl":"https://engineering.fb.com/wp-content/uploads/2026/06/Privacy-Aware-Infrastructure-in-the-AI-Native-Era-HERO.png?w=580&h=326&crop=1","alt":"","afterParagraph":26,"url":"/media/articles/cmsgi5rk106bero5qx58elyce/be68a6747bece030.webp"},{"sourceUrl":"https://engineering.fb.com/wp-content/uploads/2026/05/SilverTorch-Hero-Final.png?w=580&h=326&crop=1","alt":"","afterParagraph":26,"url":"/media/articles/cmsgi5rk106bero5qx58elyce/4e3461f0bceaaf98.webp"}],"mediaStatus":"ok","articleBodyZh":["每天，Meta 的推荐平台处理数十亿次用户互动，生成丰富的时间信号，捕捉用户在产品、广告和内容中的个性化偏好和意图。在我们 2024 年关于广告推荐序列学习的文章中：https://engineering.fb.com/2024/11/19/data-infrastructure/sequence-learning-personalized-ads-recommendations/，我们展示了如何通过建模用户动作的顺序和时间（而不是依赖静态的、手动设计的稀疏特征）来生成更丰富、能够感知序列的用户兴趣和广告偏好表示。","本文更进一步，介绍了两个架构突破，使我们能够将序列学习的进展从基础创新扩展到具有可预测、类大语言模型扩展规律的生产平台：(1) 一个多阶段序列模型，将繁重的离线用户建模与轻量的在线排序任务解耦；(2) 一种基于密集标记化和目标感知注意力的学习技术，可以直接从数据中高效学习特征交互。","结合我们更广泛的模型创新，这些进展在 Instagram 上带来了累计 6% 的转化率提升，在 Facebook 上带来了 3% 的转化率提升，以及 Facebook 广告点击率的 3.5% 提升。这个统一的序列建模平台是 Meta 生成式广告推荐模型（GEM）的核心组件：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/，帮助利用这一学习范式对用户行为的全面理解，以最大化对广告主的价值。","广告推荐系统必须在毫秒级别内检索和排序数千条广告，每秒处理数百万个候选项。为了应对这一规模，一些序列模型方法依赖混合模型配置，其中一个特定模型处理用户事件序列，另一个模型处理稀疏特征交互。","虽然这种混合方法在满足生产需求方面有效，但存在潜在的权衡：","同时扩展时间序列长度和处理这些序列的变换器模型，可能会将混合方法的权衡变成瓶颈，从而限制提升用户广告体验和广告客户活动绩效的能力。","我们在序列学习中取得了两个根本性的架构突破，解决了模型复杂性与服务效率之间的核心矛盾：（1）一种多阶段序列模型，将离线用户建模与在线排序解耦；（2）具有目标感知注意力学习范式的密集分词。二者结合提供了一种灵活的生产策略，有助于序列学习模型的泛化，并建立了类似大语言模型（LLM）的扩展规律，可预测地平衡模型性能与计算需求。","为了解决扩展效率问题，我们开发了一种多阶段模型，使基于变换器的序列模型能够以高效的计算方式进行扩展。将序列模型分为两个互补阶段（上游/离线用户建模和下游/在线排序），使模型容量能够扩展，从而在性能不断提升的同时，不需要成比例增加服务资源。","在图 1 中，左侧面板显示了离线用户模型。它异步处理长用户历史记录，并生成捕捉深层行为模式的缓存嵌入。右侧面板显示了在线排序模型，它将这些缓存表示与实时广告候选信号结合，生成最终排序。两个阶段之间的箭头表示从离线 → 在线排序模型传递的用户特征嵌入。","用户端特征通过深度变换器上游模型异步处理。这些模型可以扩展到数千序列长度的多个变换器层，并生成在用户级别预计算和缓存的嵌入。上游模型严格将用户特征与广告和上下文特征分离，以确保用户嵌入不依赖于任何特定广告候选。","离线用户模型的表示与在线排序模型互为补充，在线排序模型使用最新的用户信号和广告候选信息进行实时排序。此阶段针对速度进行了优化，满足严格的延迟要求，同时利用离线计算的深度表示。","将序列建模系统分为两个独立但互补的阶段，使得在不增加在线排序模型服务成本的情况下，可以沿着离线用户模型的扩展曲线提高模型复杂度。","这种分词方法将稀疏特征与序列行为数据集成到一个密集词汇表中，使注意力机制能够独立发现交互。与传统依赖手工设计表示来捕捉稀疏跨特征交互的推荐系统不同，这种方法让模型直接从数据中学习这些交互。","分词后的稀疏特征和广告候选信息与用户行为序列融合，然后通过一种内存高效的多头注意力形式处理，该方法使每一层都能将用户的过去行为与正在评分的特定广告进行权衡。多层对齐的注意力块堆叠且注意力分布稳定，使每一层能够捕捉目标广告与用户历史行为之间的高阶交互，逐步将长序列提炼成紧凑表示。","在处理实际广告流量时，多阶段序列模型显示出广告推荐的可预测扩展规律，这类似于大型语言模型中观察到的规律。性能提升与计算量呈对数线性关系，相比其他基于Transformer的序列模型，扩展效率显著提高。图2通过展示计算量（FLOPs）与模型性能（以归一化熵NE衡量）之间的关系，从模型深度、内容/语义增强、模型宽度和序列长度等多个维度阐释了这些扩展特性。","与处理密集连续文本的LLM不同，广告推荐系统必须将稀疏ID特征与用户时间序列集成。尽管存在结构差异，但LLM式扩展现象的出现强烈表明该模型架构适用于进一步的序列学习应用。","我们已经确定了四个杠杆点，预计它们将有助于开启扩展规律的前沿领域：","优化性能需要在模型深度、宽度和序列长度上实现均衡增长。如果只在单一维度上进行扩展，其他维度可能会成为性能提升的瓶颈，从而导致收益递减。这与大型语言模型（LLM）扩展规律研究的发现相呼应，我们称之为扩展协同原则。","多阶段架构提供了一个可调节的杠杆，用于上下调整离线或在线模型的规模。扩展在线排序模型在计算单位内可以带来更陡的性能提升，但受限于服务/请求时间要求。扩展离线模型（如图 2 所示）则呈现更平缓的曲线，但其异步推理避免了延迟限制，允许按不受限制的速度进行扩展。","随着序列长度的增加，性能持续提升，但一个重要发现是序列多样性优于序列同质性。多样化的动作类型组合（例如浏览、点击、转化）比单一动作类型的序列效果更佳。这一发现表明，混合多样的参与类型和广泛的时间覆盖能比单独的高信号动作同质序列提供更丰富的用户行为表示。","基础模型的语义内容特征可以补充传统协同过滤（即哪些用户与哪些物品互动）信号。它们在冷启动场景（例如新的广告或广告主历史互动数据有限）中尤其有帮助。通过解决推荐系统的这一持续挑战，我们提高了信号覆盖，从而应对推荐系统中基本的稀疏问题。","多阶段序列建模架构在三个维度上产生了影响：","通过建模数千条用户事件序列（如点击、浏览和购买），离线模型生成了高度细化的用户表示。这种深度的行为理解提升了 Meta 系列应用的广告相关性和转化率。结合我们更广泛的建模创新，这些基于序列的表示在 Instagram 上带来了总体转化率提升 6%，在 Facebook 上提升了 3%，以及 Facebook 广告点击提升了 3.5%。","与混合方法相比，这种两阶段设计在计算效率更高的情况下提供了性能提升。初步评估显示广告排名质量有所提高，对服务资源的影响最小，这证实了模型复杂性和生产效率可以同时扩展。","作为 GEM 的核心部分：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/，这种用于序列学习的模型架构被设计为具有泛化能力，其中相同的多阶段骨干网络和可扩展特性可以以最小的调整和开销扩展到任何广告排序任务。","序列模型的扩展规律显示尚未出现饱和迹象。在实现架构等效后，模型复杂性扩展可以借鉴在大型语言模型领域验证过的技术（例如专家混合、跨用户计算共享、先进的注意力机制），有可能在最佳性能/效率权衡下持续扩展。","关于该模型架构及其扩展特性的详细技术出版物可参见我们的论文：“LLaTTE：大规模广告推荐多阶段序列建模的扩展规律”：https://arxiv.org/pdf/2601.20083。","Meta 相信通过开源技术建设社区。探索我们在人工智能、数据基础设施、开发工具、前端、编程语言、平台、安全、虚拟现实等领域的最新项目。","Meta 的工程博客是技术新闻资源，为对我们如何解决大型技术挑战感兴趣的工程师提供信息。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Meta 介绍多阶段序列模型，将高计算量的离线用户建模与轻量在线排序分离，并结合密集 token 化和目标感知注意力，以可预测的扩展规律提升广告排序。","background":"Meta 表示，其推荐平台每天处理数十亿次用户交互，广告系统还需在毫秒内检索和排序数千条广告。既有混合架构分别处理用户事件序列与稀疏特征交互，但扩大序列长度和 Transformer 规模可能形成瓶颈。","viewpoint":"Aioga 判断，这项更新的重点不只是扩大模型，而是通过离线与在线分阶段设计缓解模型复杂度和服务效率之间的矛盾。所谓 LLM 式扩展定律，材料将其描述为性能与计算之间可预测的平衡。","implications":"Meta 称相关进展与其他模型创新共同带来 Instagram 转化率累计提升 6%，Facebook 转化率提升 3%、广告点击量提升 3.5%。值得关注的是，这些结果属于多项创新的累计贡献，不能仅归因于单一架构。","nextStep":"值得关注的是，Meta 后续是否披露扩展规律的测量方法、计算成本、线上延迟及不同阶段的独立增益。现有材料确认该平台是 GEM 的核心组件，但未提供各项改动的单独归因数据。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-08-05T20:24:55.832Z","sourceHash":"830122342ebee6f0","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["候选内容准确区分了 Meta 的原始表述、Aioga 的判断以及材料未披露的信息。","对效果数据增加了“多项创新的累计贡献、不能仅归因于单一架构”的限定，符合来源材料，避免了误导性归因。","“LLM 式扩展定律”被解释为性能与计算之间的可预测平衡，与正文摘录一致。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["产品更新","Meta Engineering Blog（RSS）"],"translations":{"zh-CN":{"title":"Meta 广告排序的多阶段序列模型：从用户序列到 LLM 式扩展定律","summary":"Meta 推出多阶段序列模型，将离线用户建模与在线排序解耦，并采用密集 token 化与目标感知注意力，使序列学习具备可预测的 LLM 式扩展定律。该架构已为 Instagram 转化率带来 6% 的累计提升，Facebook 转化率提升 3%、广告点击量提升 3.5%，并成为 Meta 生成式广告推荐模型（GEM）的核心组件。","category":"产品更新","source":"engineering.fb.com","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta 广告排序的多阶段序列模型：从用户序列到 LLM 式扩展定律 - Aioga AI资讯","description":"Meta 推出多阶段序列模型，将离线用户建模与在线排序解耦，并采用密集 token 化与目标感知注意力，使序列学习具备可预测的 LLM 式扩展定律。该架构已为 Instagram 转化率带来 6% 的累计提升，Facebook 转化率提升 3%、广告点击量提升 3.5%，并成为 Meta 生成式广告推荐模型（GEM）的核心组件。","url":"https://www.aioga.com/news/cmsgi5rk106bero5qx58elyce/","articleBody":["每天，Meta 的推荐平台处理数十亿次用户互动，生成丰富的时间信号，捕捉用户在产品、广告和内容中的个性化偏好和意图。在我们 2024 年关于广告推荐序列学习的文章中：https://engineering.fb.com/2024/11/19/data-infrastructure/sequence-learning-personalized-ads-recommendations/，我们展示了如何通过建模用户动作的顺序和时间（而不是依赖静态的、手动设计的稀疏特征）来生成更丰富、能够感知序列的用户兴趣和广告偏好表示。","本文更进一步，介绍了两个架构突破，使我们能够将序列学习的进展从基础创新扩展到具有可预测、类大语言模型扩展规律的生产平台：(1) 一个多阶段序列模型，将繁重的离线用户建模与轻量的在线排序任务解耦；(2) 一种基于密集标记化和目标感知注意力的学习技术，可以直接从数据中高效学习特征交互。","结合我们更广泛的模型创新，这些进展在 Instagram 上带来了累计 6% 的转化率提升，在 Facebook 上带来了 3% 的转化率提升，以及 Facebook 广告点击率的 3.5% 提升。这个统一的序列建模平台是 Meta 生成式广告推荐模型（GEM）的核心组件：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/，帮助利用这一学习范式对用户行为的全面理解，以最大化对广告主的价值。","广告推荐系统必须在毫秒级别内检索和排序数千条广告，每秒处理数百万个候选项。为了应对这一规模，一些序列模型方法依赖混合模型配置，其中一个特定模型处理用户事件序列，另一个模型处理稀疏特征交互。","虽然这种混合方法在满足生产需求方面有效，但存在潜在的权衡：","同时扩展时间序列长度和处理这些序列的变换器模型，可能会将混合方法的权衡变成瓶颈，从而限制提升用户广告体验和广告客户活动绩效的能力。","我们在序列学习中取得了两个根本性的架构突破，解决了模型复杂性与服务效率之间的核心矛盾：（1）一种多阶段序列模型，将离线用户建模与在线排序解耦；（2）具有目标感知注意力学习范式的密集分词。二者结合提供了一种灵活的生产策略，有助于序列学习模型的泛化，并建立了类似大语言模型（LLM）的扩展规律，可预测地平衡模型性能与计算需求。","为了解决扩展效率问题，我们开发了一种多阶段模型，使基于变换器的序列模型能够以高效的计算方式进行扩展。将序列模型分为两个互补阶段（上游/离线用户建模和下游/在线排序），使模型容量能够扩展，从而在性能不断提升的同时，不需要成比例增加服务资源。","在图 1 中，左侧面板显示了离线用户模型。它异步处理长用户历史记录，并生成捕捉深层行为模式的缓存嵌入。右侧面板显示了在线排序模型，它将这些缓存表示与实时广告候选信号结合，生成最终排序。两个阶段之间的箭头表示从离线 → 在线排序模型传递的用户特征嵌入。","用户端特征通过深度变换器上游模型异步处理。这些模型可以扩展到数千序列长度的多个变换器层，并生成在用户级别预计算和缓存的嵌入。上游模型严格将用户特征与广告和上下文特征分离，以确保用户嵌入不依赖于任何特定广告候选。","离线用户模型的表示与在线排序模型互为补充，在线排序模型使用最新的用户信号和广告候选信息进行实时排序。此阶段针对速度进行了优化，满足严格的延迟要求，同时利用离线计算的深度表示。","将序列建模系统分为两个独立但互补的阶段，使得在不增加在线排序模型服务成本的情况下，可以沿着离线用户模型的扩展曲线提高模型复杂度。","这种分词方法将稀疏特征与序列行为数据集成到一个密集词汇表中，使注意力机制能够独立发现交互。与传统依赖手工设计表示来捕捉稀疏跨特征交互的推荐系统不同，这种方法让模型直接从数据中学习这些交互。","分词后的稀疏特征和广告候选信息与用户行为序列融合，然后通过一种内存高效的多头注意力形式处理，该方法使每一层都能将用户的过去行为与正在评分的特定广告进行权衡。多层对齐的注意力块堆叠且注意力分布稳定，使每一层能够捕捉目标广告与用户历史行为之间的高阶交互，逐步将长序列提炼成紧凑表示。","在处理实际广告流量时，多阶段序列模型显示出广告推荐的可预测扩展规律，这类似于大型语言模型中观察到的规律。性能提升与计算量呈对数线性关系，相比其他基于Transformer的序列模型，扩展效率显著提高。图2通过展示计算量（FLOPs）与模型性能（以归一化熵NE衡量）之间的关系，从模型深度、内容/语义增强、模型宽度和序列长度等多个维度阐释了这些扩展特性。","与处理密集连续文本的LLM不同，广告推荐系统必须将稀疏ID特征与用户时间序列集成。尽管存在结构差异，但LLM式扩展现象的出现强烈表明该模型架构适用于进一步的序列学习应用。","我们已经确定了四个杠杆点，预计它们将有助于开启扩展规律的前沿领域：","优化性能需要在模型深度、宽度和序列长度上实现均衡增长。如果只在单一维度上进行扩展，其他维度可能会成为性能提升的瓶颈，从而导致收益递减。这与大型语言模型（LLM）扩展规律研究的发现相呼应，我们称之为扩展协同原则。","多阶段架构提供了一个可调节的杠杆，用于上下调整离线或在线模型的规模。扩展在线排序模型在计算单位内可以带来更陡的性能提升，但受限于服务/请求时间要求。扩展离线模型（如图 2 所示）则呈现更平缓的曲线，但其异步推理避免了延迟限制，允许按不受限制的速度进行扩展。","随着序列长度的增加，性能持续提升，但一个重要发现是序列多样性优于序列同质性。多样化的动作类型组合（例如浏览、点击、转化）比单一动作类型的序列效果更佳。这一发现表明，混合多样的参与类型和广泛的时间覆盖能比单独的高信号动作同质序列提供更丰富的用户行为表示。","基础模型的语义内容特征可以补充传统协同过滤（即哪些用户与哪些物品互动）信号。它们在冷启动场景（例如新的广告或广告主历史互动数据有限）中尤其有帮助。通过解决推荐系统的这一持续挑战，我们提高了信号覆盖，从而应对推荐系统中基本的稀疏问题。","多阶段序列建模架构在三个维度上产生了影响：","通过建模数千条用户事件序列（如点击、浏览和购买），离线模型生成了高度细化的用户表示。这种深度的行为理解提升了 Meta 系列应用的广告相关性和转化率。结合我们更广泛的建模创新，这些基于序列的表示在 Instagram 上带来了总体转化率提升 6%，在 Facebook 上提升了 3%，以及 Facebook 广告点击提升了 3.5%。","与混合方法相比，这种两阶段设计在计算效率更高的情况下提供了性能提升。初步评估显示广告排名质量有所提高，对服务资源的影响最小，这证实了模型复杂性和生产效率可以同时扩展。","作为 GEM 的核心部分：https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/，这种用于序列学习的模型架构被设计为具有泛化能力，其中相同的多阶段骨干网络和可扩展特性可以以最小的调整和开销扩展到任何广告排序任务。","序列模型的扩展规律显示尚未出现饱和迹象。在实现架构等效后，模型复杂性扩展可以借鉴在大型语言模型领域验证过的技术（例如专家混合、跨用户计算共享、先进的注意力机制），有可能在最佳性能/效率权衡下持续扩展。","关于该模型架构及其扩展特性的详细技术出版物可参见我们的论文：“LLaTTE：大规模广告推荐多阶段序列建模的扩展规律”：https://arxiv.org/pdf/2601.20083。","Meta 相信通过开源技术建设社区。探索我们在人工智能、数据基础设施、开发工具、前端、编程语言、平台、安全、虚拟现实等领域的最新项目。","Meta 的工程博客是技术新闻资源，为对我们如何解决大型技术挑战感兴趣的工程师提供信息。"]},"en":{"title":"Meta's Multi-Stage Sequence Model for Ad Ranking: From User Sequences to LLM-Style Scaling Laws","summary":"Meta has launched a multi-stage sequential model that decouples offline user modeling from online ranking, and employs dense tokenization and goal-aware attention, enabling sequence learning to have predictable LLM-style scaling laws. This architecture has brought a cumulative 6% increase in Instagram conversion rates, a 3% increase in Facebook conversion rates, and a 3.5% increase in ad clicks, and has become a core component of Meta's Generative Ad Model (GEM).","category":"Products","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta's Multi-Stage Sequence Model for Ad Ranking: From User Sequences to LLM-Style Scaling Laws - Aioga AI News","description":"Meta has launched a multi-stage sequential model that decouples offline user modeling from online ranking, and employs dense tokenization and goal-aware attention, enabling sequenc...","url":"https://www.aioga.com/en/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:01:49.041Z"},"ja":{"title":"Meta 広告のランキングにおける多段階シーケンスモデル：ユーザーシーケンスからLLM式の拡張法則まで","summary":"Metaは多段階シーケンスモデルを導入し、オフラインユーザーのモデリングとオンラインのランキングを分離し、密なトークン化とターゲット認識注意機構を採用することで、シーケンス学習に予測可能なLLM式スケーリング則を持たせています。このアーキテクチャにより、Instagramのコンバージョン率は累積で6％向上し、Facebookのコンバージョン率は3％増加、広告クリック数は3.5％増加し、Metaの生成型広告推薦モデル（GEM）のコアコンポーネントとなっています。","category":"製品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta 広告のランキングにおける多段階シーケンスモデル：ユーザーシーケンスからLLM式の拡張法則まで - Aioga AIニュース","description":"Metaは多段階シーケンスモデルを導入し、オフラインユーザーのモデリングとオンラインのランキングを分離し、密なトークン化とターゲット認識注意機構を採用することで、シーケンス学習に予測可能なLLM式スケーリング則を持たせています。このアーキテクチャにより、Instagramのコンバージョン率は累積で6％向上し、Facebookのコンバージョン率は3％増加、広告...","url":"https://www.aioga.com/ja/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:01:58.951Z"},"ko":{"title":"Meta 광고 순위의 다단계 시퀀스 모델: 사용자 시퀀스에서 LLM식 확장 법칙까지","summary":"Meta는 다단계 시퀀스 모델을 출시하여 오프라인 사용자 모델링과 온라인 순위를 분리하고, 밀집 토큰화와 목표 인식 주의를 사용하여 시퀀스 학습이 예측 가능한 LLM 스타일 확장 법칙을 갖도록 만들었습니다. 이 아키텍처는 Instagram 전환율에 누적 6% 향상을 가져왔고, Facebook 전환율 3%, 광고 클릭 수 3.5% 증가를 이루었으며, Meta 생성형 광고 추천 모델(GEM)의 핵심 구성 요소가 되었습니다.","category":"제품 업데이트","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta 광고 순위의 다단계 시퀀스 모델: 사용자 시퀀스에서 LLM식 확장 법칙까지 - Aioga AI 뉴스","description":"Meta는 다단계 시퀀스 모델을 출시하여 오프라인 사용자 모델링과 온라인 순위를 분리하고, 밀집 토큰화와 목표 인식 주의를 사용하여 시퀀스 학습이 예측 가능한 LLM 스타일 확장 법칙을 갖도록 만들었습니다. 이 아키텍처는 Instagram 전환율에 누적 6% 향상을 가져왔고, Facebook 전환율 3%, 광고 클릭 수...","url":"https://www.aioga.com/ko/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:02:34.559Z"},"es":{"title":"Modelo de secuencia multi-etapa para la clasificación de anuncios de Meta: de la secuencia de usuarios a la ley de escalamiento estilo LLM","summary":"Meta ha lanzado un modelo de secuencias en varias etapas que desacopla el modelado offline de usuarios del ordenamiento en línea, empleando una tokenización densa y atención consciente de objetos, permitiendo que el aprendizaje de secuencias siga leyes de escalado predecibles al estilo LLM. Esta arquitectura ha provocado un aumento acumulado del 6% en las tasas de conversión en Instagram, un aumento del 3% en Facebook y un incremento del 3,5% en los clics publicitarios, convirtiéndose en un componente central del modelo de recomendación generativa de anuncios (GEM) de Meta.","category":"Productos","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Modelo de secuencia multi-etapa para la clasificación de anuncios de Meta: de la secuencia de usuarios a la ley de escalamiento estilo LLM - Aioga Noticias de IA","description":"Meta ha lanzado un modelo de secuencias en varias etapas que desacopla el modelado offline de usuarios del ordenamiento en línea, empleando una tokenización densa y atención consci...","url":"https://www.aioga.com/es/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:02:36.505Z"},"fr":{"title":"Modèle séquentiel à plusieurs étapes pour le classement des publicités Meta : des séquences utilisateur à la loi d'expansion de type LLM","summary":"Meta a lancé un modèle de séquence multi-étapes, séparant la modélisation des utilisateurs hors ligne du classement en ligne, et utilisant une tokenisation dense ainsi qu'une attention consciente des objectifs, permettant à l'apprentissage séquentiel de posséder une loi d'expansion prévisible de type LLM. Cette architecture a déjà apporté une augmentation cumulative de 6 % du taux de conversion sur Instagram, 3 % sur Facebook, une hausse de 3,5 % des clics publicitaires, et est devenue un composant clé du modèle de recommandation publicitaire génératif de Meta (GEM).","category":"Produits","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Modèle séquentiel à plusieurs étapes pour le classement des publicités Meta : des séquences utilisateur à la loi d'expansion de type LLM - Aioga Actualités IA","description":"Meta a lancé un modèle de séquence multi-étapes, séparant la modélisation des utilisateurs hors ligne du classement en ligne, et utilisant une tokenisation dense ainsi qu'une atten...","url":"https://www.aioga.com/fr/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:03:11.299Z"},"de":{"title":"Mehrstufiges Sequenzmodell zur Anzeigenplatzierung bei Meta: Von Benutzersequenzen bis zu LLM-ähnlichen Skalierungsgesetzen","summary":"Meta hat ein mehrstufiges Sequenzmodell eingeführt, das die Offline-Nutzer-Modellierung von der Online-Rangordnung entkoppelt und dabei dichte Tokenisierung sowie zielgerichtete Aufmerksamkeit verwendet, sodass das Sequenzlernen eine vorhersehbare LLM-ähnliche Skalierungsgesetzmäßigkeit besitzt. Diese Architektur hat bereits die Konversionsrate von Instagram um kumulativ 6 % gesteigert, die Konversionsrate von Facebook um 3 % erhöht und die Werbeklicks um 3,5 % gesteigert und ist die Kernkomponente des generativen Werbeempfehlungsmodells (GEM) von Meta geworden.","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Mehrstufiges Sequenzmodell zur Anzeigenplatzierung bei Meta: Von Benutzersequenzen bis zu LLM-ähnlichen Skalierungsgesetzen - Aioga KI-News","description":"Meta hat ein mehrstufiges Sequenzmodell eingeführt, das die Offline-Nutzer-Modellierung von der Online-Rangordnung entkoppelt und dabei dichte Tokenisierung sowie zielgerichtete Au...","url":"https://www.aioga.com/de/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:03:16.743Z"},"pt-BR":{"title":"Modelo de sequência em múltiplas etapas de classificação de anúncios da Meta: da sequência do usuário à lei de expansão estilo LLM","summary":"A Meta lançou um modelo de sequência em múltiplos estágios, que separa a modelagem de usuários offline da classificação online, e utiliza tokenização densa e atenção sensível a objetivos, tornando o aprendizado de sequência capaz de possuir uma lei de escalabilidade previsível no estilo LLM. Essa arquitetura já trouxe um aumento acumulado de 6% na taxa de conversão do Instagram, 3% na taxa de conversão do Facebook, 3,5% no número de cliques em anúncios, e se tornou um componente central do modelo de recomendação de anúncios generativos (GEM) da Meta.","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Modelo de sequência em múltiplas etapas de classificação de anúncios da Meta: da sequência do usuário à lei de expansão estilo LLM - Aioga Notícias de IA","description":"A Meta lançou um modelo de sequência em múltiplos estágios, que separa a modelagem de usuários offline da classificação online, e utiliza tokenização densa e atenção sensível a obj...","url":"https://www.aioga.com/pt-BR/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:03:50.849Z"},"ru":{"title":"Многоступенчатая последовательная модель ранжирования рекламы Meta: от пользовательских последовательностей к закону масштабирования в стиле LLM","summary":"Meta представила многоэтапную модель последовательностей, которая разъединяет моделирование оффлайн-пользователей и онлайн-ранжирование, а также использует плотную токенизацию и целевую внимательность, позволяя обучению последовательностей обладать предсказуемым законом масштабирования в стиле LLM. Эта архитектура уже принесла Instagram 6% накопленного прироста конверсии, Facebook — 3% прироста конверсии и 3,5% прироста кликов по рекламе, и стала ключевым компонентом модели генеративной рекламной рекомендации Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Многоступенчатая последовательная модель ранжирования рекламы Meta: от пользовательских последовательностей к закону масштабирования в стиле LLM - Aioga Новости ИИ","description":"Meta представила многоэтапную модель последовательностей, которая разъединяет моделирование оффлайн-пользователей и онлайн-ранжирование, а также использует плотную токенизацию и це...","url":"https://www.aioga.com/ru/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:03:53.360Z"},"ar":{"title":"نموذج التسلسل متعدد المراحل لترتيب الإعلانات في ميتا: من تسلسل المستخدم إلى قانون التوسع على نمط LLM","summary":"أطلقت شركة Meta نموذج تسلسل متعدد المراحل، يفصل بين نمذجة المستخدمين خارج الإنترنت والترتيب عبر الإنترنت، ويستخدم الترميز المكثف للرموز والانتباه الواعي بالهدف، مما يجعل تعلم التسلسل يمتلك قانون توسع متوقع على غرار LLM. وقد أدى هذا الهيكل إلى زيادة تراكمية بنسبة 6٪ في معدل التحويل على إنستغرام، وزيادة بنسبة 3٪ في معدل التحويل على فيسبوك، وزيادة بنسبة 3.5٪ في نقرات الإعلانات، وأصبح المكون الأساسي لنموذج توصية الإعلانات التوليدية لشركة Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"نموذج التسلسل متعدد المراحل لترتيب الإعلانات في ميتا: من تسلسل المستخدم إلى قانون التوسع على نمط LLM - Aioga أخبار الذكاء الاصطناعي","description":"أطلقت شركة Meta نموذج تسلسل متعدد المراحل، يفصل بين نمذجة المستخدمين خارج الإنترنت والترتيب عبر الإنترنت، ويستخدم الترميز المكثف للرموز والانتباه الواعي بالهدف، مما يجعل تعلم التسل...","url":"https://www.aioga.com/ar/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:04:34.637Z"},"hi":{"title":"मेटा विज्ञापन क्रमांकन का बहु-चरणीय अनुक्रम मॉडल: उपयोगकर्ता अनुक्रम से LLM शैली विस्तार नियम तक","summary":"Meta ने बहु-चरणीय अनुक्रम मॉडल लॉन्च किया, जो ऑफलाइन उपयोगकर्ता मॉडलिंग और ऑनलाइन रैंकिंग को अलग करता है, और घनत्व टोकनिंग और लक्ष्य-संवेदनशील ध्यान का उपयोग करता है, जिससे अनुक्रम सीखने में पूर्वानुमेय LLM-शैली का विस्तार कानून प्राप्त होता है। इस आर्किटेक्चर ने Instagram रूपांतरण दर में कुल 6% की वृद्धि, Facebook रूपांतरण दर में 3% की वृद्धि, विज्ञापन क्लिक दर में 3.5% की वृद्धि लायी है, और Meta के जनरेटिव विज्ञापन सिफारिश मॉडल (GEM) का मुख्य घटक बन गया है।","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"मेटा विज्ञापन क्रमांकन का बहु-चरणीय अनुक्रम मॉडल: उपयोगकर्ता अनुक्रम से LLM शैली विस्तार नियम तक - Aioga AI समाचार","description":"Meta ने बहु-चरणीय अनुक्रम मॉडल लॉन्च किया, जो ऑफलाइन उपयोगकर्ता मॉडलिंग और ऑनलाइन रैंकिंग को अलग करता है, और घनत्व टोकनिंग और लक्ष्य-संवेदनशील ध्यान का उपयोग करता है, जिससे अनुक्रम...","url":"https://www.aioga.com/hi/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:04:31.760Z"},"it":{"title":"Modello a sequenze multi-fase per il ranking degli annunci di Meta: dalla sequenza degli utenti alla legge di scaling in stile LLM","summary":"Meta ha lanciato un modello sequenziale a più fasi, che scinde la modellazione degli utenti offline dal ranking online, e utilizza una tokenizzazione densa e un'attenzione sensibile agli obiettivi, permettendo all'apprendimento sequenziale di avere leggi di scalabilità prevedibili in stile LLM. Questa architettura ha portato a un aumento cumulativo del 6% del tasso di conversione su Instagram, del 3% su Facebook e del 3,5% dei clic sugli annunci, ed è diventata il componente centrale del modello di raccomandazione pubblicitaria generativa di Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Modello a sequenze multi-fase per il ranking degli annunci di Meta: dalla sequenza degli utenti alla legge di scaling in stile LLM - Aioga Notizie IA","description":"Meta ha lanciato un modello sequenziale a più fasi, che scinde la modellazione degli utenti offline dal ranking online, e utilizza una tokenizzazione densa e un'attenzione sensibil...","url":"https://www.aioga.com/it/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:05:15.950Z"},"nl":{"title":"Meta-advertentie rangschikking multi-fasen sequentiemodel: van gebruikerssequentie naar LLM-achtige uitbreidingswet","summary":"Meta heeft een multi-stage sequentiemodel geïntroduceerd, dat offline gebruikersmodellering en online rangschikking ontkoppelt, en gebruikmaakt van dicht tokenisatie en doelbewuste aandacht, waardoor sequentieel leren een voorspelbare LLM-achtige schaalwet krijgt. Deze architectuur heeft geleid tot een cumulatieve stijging van 6% in de conversieratio van Instagram, een stijging van 3% in de conversieratio van Facebook en een stijging van 3,5% in advertentieklikken, en is het kernonderdeel geworden van Meta's generatieve advertentie-aanbevelingsmodel (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta-advertentie rangschikking multi-fasen sequentiemodel: van gebruikerssequentie naar LLM-achtige uitbreidingswet - Aioga AI-nieuws","description":"Meta heeft een multi-stage sequentiemodel geïntroduceerd, dat offline gebruikersmodellering en online rangschikking ontkoppelt, en gebruikmaakt van dicht tokenisatie en doelbewuste...","url":"https://www.aioga.com/nl/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:05:10.231Z"},"tr":{"title":"Meta reklam sıralamasının çok aşamalı sıra modeli: Kullanıcı sırasından LLM tarzı genişleme yasasına","summary":"Meta, çevrimdışı kullanıcı modellemesini çevrimiçi sıralamadan ayıran çok aşamalı bir sıra modeli geliştirdi ve yoğun tokenlaştırma ile hedef algılayıcı dikkat kullanarak sıra öğreniminin öngörülebilir LLM tarzı ölçeklenme yasasına sahip olmasını sağladı. Bu mimari, Instagram dönüşüm oranında %6'lık kümülatif bir artış, Facebook dönüşüm oranında %3 ve reklam tıklanma oranında %3,5 artış sağladı ve Meta'nın üretken reklam öneri modeli (GEM) için temel bir bileşen haline geldi.","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Meta reklam sıralamasının çok aşamalı sıra modeli: Kullanıcı sırasından LLM tarzı genişleme yasasına - Aioga AI Haberleri","description":"Meta, çevrimdışı kullanıcı modellemesini çevrimiçi sıralamadan ayıran çok aşamalı bir sıra modeli geliştirdi ve yoğun tokenlaştırma ile hedef algılayıcı dikkat kullanarak sıra öğre...","url":"https://www.aioga.com/tr/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:05:48.562Z"},"vi":{"title":"Mô hình tuần tự đa giai đoạn của xếp hạng quảng cáo Meta: Từ trình tự người dùng đến quy luật mở rộng kiểu LLM","summary":"Meta ra mắt mô hình chuỗi nhiều giai đoạn, tách biệt việc mô hình hóa người dùng ngoại tuyến và xếp hạng trực tuyến, đồng thời sử dụng mã hóa token dày đặc và chú ý nhận thức mục tiêu, giúp học chuỗi có được quy luật mở rộng kiểu LLM có thể dự đoán. Kiến trúc này đã mang lại tăng trưởng tích lũy 6% về tỷ lệ chuyển đổi cho Instagram, tăng 3% tỷ lệ chuyển đổi trên Facebook, tăng 3,5% lượt nhấp quảng cáo, và trở thành thành phần cốt lõi của mô hình đề xuất quảng cáo tạo sinh của Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Mô hình tuần tự đa giai đoạn của xếp hạng quảng cáo Meta: Từ trình tự người dùng đến quy luật mở rộng kiểu LLM - Tin tức AI Aioga","description":"Meta ra mắt mô hình chuỗi nhiều giai đoạn, tách biệt việc mô hình hóa người dùng ngoại tuyến và xếp hạng trực tuyến, đồng thời sử dụng mã hóa token dày đặc và chú ý nhận thức mục t...","url":"https://www.aioga.com/vi/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:05:47.836Z"},"id":{"title":"Model Urutan Multi-Tahap untuk Penyortiran Iklan Meta: Dari Urutan Pengguna hingga Hukum Perluasan ala LLM","summary":"Meta meluncurkan model urutan multi-tahap, memisahkan pemodelan pengguna offline dari pengurutan online, dan menggunakan tokenisasi padat serta perhatian yang peka terhadap tujuan, sehingga pembelajaran urutan memiliki hukum skala mirip LLM yang dapat diprediksi. Arsitektur ini telah memberikan peningkatan kumulatif 6% pada rasio konversi Instagram, peningkatan 3% pada rasio konversi Facebook, peningkatan 3,5% pada klik iklan, dan menjadi komponen inti dari model rekomendasi iklan generatif Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Model Urutan Multi-Tahap untuk Penyortiran Iklan Meta: Dari Urutan Pengguna hingga Hukum Perluasan ala LLM - Berita AI Aioga","description":"Meta meluncurkan model urutan multi-tahap, memisahkan pemodelan pengguna offline dari pengurutan online, dan menggunakan tokenisasi padat serta perhatian yang peka terhadap tujuan,...","url":"https://www.aioga.com/id/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:06:23.111Z"},"th":{"title":"โมเดลลำดับหลายขั้นตอนของการจัดอันดับโฆษณา Meta: จากลำดับผู้ใช้ไปสู่กฎการขยายแบบ LLM","summary":"Meta เปิดตัวโมเดลลำดับหลายขั้นตอน ซึ่งแยกการสร้างแบบจำลองผู้ใช้ออฟไลน์ออกจากการจัดลำดับออนไลน์ และใช้การทำ token หนาแน่นพร้อมความสนใจที่รับรู้เป้าหมาย ทำให้การเรียนรู้ลำดับมีแนวโน้มการขยายแบบ LLM ที่สามารถคาดการณ์ได้ สถาปัตยกรรมนี้ได้เพิ่มอัตราการแปลงของ Instagram รวม 6% เพิ่มอัตราการแปลงของ Facebook 3% เพิ่มการคลิกโฆษณา 3.5% และกลายเป็นส่วนประกอบหลักของโมเดลการแนะนำโฆษณาสร้างสรรค์ของ Meta (GEM)","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"โมเดลลำดับหลายขั้นตอนของการจัดอันดับโฆษณา Meta: จากลำดับผู้ใช้ไปสู่กฎการขยายแบบ LLM - ข่าว AI Aioga","description":"Meta เปิดตัวโมเดลลำดับหลายขั้นตอน ซึ่งแยกการสร้างแบบจำลองผู้ใช้ออฟไลน์ออกจากการจัดลำดับออนไลน์ และใช้การทำ token หนาแน่นพร้อมความสนใจที่รับรู้เป้าหมาย ทำให้การเรียนรู้ลำดับมีแนวโน้...","url":"https://www.aioga.com/th/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:06:39.299Z"},"pl":{"title":"Wieloetapowy model sekwencyjny sortowania reklam Meta: od sekwencji użytkowników do prawa rozszerzania w stylu LLM","summary":"Meta wprowadziło wieloetapowy model sekwencyjny, który oddziela modelowanie użytkowników offline od rankingu online, stosując intensywną tokenizację oraz uwagę ukierunkowaną na cel, dzięki czemu nauka sekwencji ma przewidywalne prawa skalowania w stylu LLM. Ta architektura przyniosła łączny wzrost współczynnika konwersji na Instagramie o 6%, współczynnika konwersji na Facebooku o 3%, wzrost liczby kliknięć w reklamy o 3,5% i stała się kluczowym komponentem modeli rekomendacji reklam generatywnych Meta (GEM).","category":"产品更新","source":"Meta Engineering Blog（RSS）","aggregationSource":"Meta Engineering Blog（RSS）","pageTitle":"Wieloetapowy model sekwencyjny sortowania reklam Meta: od sekwencji użytkowników do prawa rozszerzania w stylu LLM - Aioga Wiadomości AI","description":"Meta wprowadziło wieloetapowy model sekwencyjny, który oddziela modelowanie użytkowników offline od rankingu online, stosując intensywną tokenizację oraz uwagę ukierunkowaną na cel...","url":"https://www.aioga.com/pl/news/cmsgi5rk106bero5qx58elyce/","contentTranslated":true,"sourceHash":"f236f5debd21c86c","translatedAt":"2026-08-05T20:07:14.729Z"}}}}