OneModel: A Unified Foundation for Platform-Scale Multi-Scenario Ranking
OneModel unifies multi-scenario ranking using long-context user modeling, achieving +0.33% Time Spent and +8.18% CTR online.
Key Findings
Methodology
OneModel employs scenario-aware feature projection and structural context encoding to map heterogeneous behaviors into shared event sequences. It builds an action-oriented long-sequence backbone with Transformer decoders, integrating Scenario-aware Information Modulation (SAIM) for cross-scenario transfer and specialization. The model supports end-to-end multi-objective training, combining self-supervised next-item prediction and supervised stream-specific tasks. Techniques like stratified user representations, feature prefetching, shared user towers, and graph-level inference optimize online serving. Core algorithms include Transformer-based causal decoders, sinusoidal positional encodings, and multi-task loss functions, enabling scalable, real-time multi-scenario ranking.
Key Results
- Deployment on Xiaohongshu's platform demonstrated offline improvements over strong baselines. Online AB tests showed a 0.33% increase in Time Spent, 1.25% in Engagement, 8.18% CTR in Feed Advertising, 3.43% lift in advertising value, and 1.19% GMV growth. The model effectively balances information sharing and scenario-specificity, reducing engineering complexity while boosting user engagement and revenue.
- Ablation studies confirmed that SAIM and stratified user representations are critical for performance gains. Multi-target training enhanced robustness across diverse tasks. Larger context lengths and model capacity correlated with better results, validating the scalability of the approach.
- The model's ability to handle heterogeneous objectives and long histories in a unified architecture demonstrates its practicality for industrial deployment, outperforming traditional isolated models in both effectiveness and efficiency.
Significance
This work addresses the long-standing challenge of integrating multiple business streams into a single, scalable recommendation framework. By leveraging shared long-sequence modeling with scene-aware modulation, it significantly improves information reuse, reduces engineering overhead, and enhances user experience. The approach paves the way for more intelligent, resource-efficient industrial recommender systems capable of supporting diverse objectives simultaneously, marking a major step forward in large-scale AI deployment. Its success at Xiaohongshu exemplifies its potential for broad industry adoption.
Technical Contribution
The paper introduces the Scenario-aware Information Modulation (SAIM) mechanism, enabling dynamic balancing of cross-scenario transfer and specialization. It designs a multi-task, end-to-end training framework with stratified user representations, combining self-supervised and supervised objectives. The architecture integrates a Transformer-based long-context backbone with explicit item-context interaction and a decoupled online serving strategy, including feature prefetching and graph-level inference. These innovations collectively advance the state-of-the-art in multi-scenario, long-sequence industrial recommendation systems.
Novelty
This is the first work to propose a unified, long-sequence generative ranking framework tailored for platform-scale multi-scenario recommendation. Unlike prior multi-task or multi-model solutions, OneModel seamlessly balances shared information and scenario-specific needs through SAIM and hierarchical user modeling. Its end-to-end training and deployment strategies address real-world industrial constraints, making it a pioneering approach in the field.
Limitations
- Despite its scalability, the model's training and inference costs remain high, especially for extremely long sequences. The reliance on large-scale data may limit applicability in cold-start or new scenarios. Further, the scene-aware modulation, while effective, could be refined for better interpretability and adaptability. Future work should focus on model compression, faster inference techniques, and adaptive scene modulation to broaden deployment scope.
Future Work
Future research will explore model compression and acceleration methods to reduce computational costs. Incorporating reinforcement learning or user feedback could dynamically optimize scene modulation strategies. Extending multi-modal data fusion, such as images and text, may enrich user representations. Additionally, developing more adaptive, interpretable scene-aware mechanisms will further enhance robustness and generalization across diverse industrial scenarios.
AI Executive Summary
In the rapidly evolving landscape of internet platforms, recommendation systems are fundamental to content delivery, user engagement, and commercial success. Traditional models, often designed for single scenarios, struggle to leverage the full spectrum of user behaviors across multiple business streams like organic content, advertising, and merchant services. These streams form a continuous, intertwined user journey, yet existing solutions typically treat them separately, leading to fragmented user representations and increased engineering complexity.
Recognizing this challenge, the authors propose OneModel, a unified, platform-scale ranking framework capable of modeling long, heterogeneous user behavior sequences across multiple scenarios. The core innovation lies in their scenario-aware feature projection and structural context encoding, which map diverse behaviors into a shared event sequence. This sequence is processed by an action-oriented backbone built on Transformer decoders, enabling scalable long-context modeling. To address cross-scenario interference, the model introduces Scenario-aware Information Modulation (SAIM), a lightweight gating mechanism that dynamically balances transfer and specialization.
The training strategy combines multi-objective learning, including self-supervised next-item prediction and supervised stream-specific tasks, with stratified user representations that fuse global preferences and short-term intents. These design choices ensure the model's robustness and adaptability in real-world deployment. Extensive offline evaluations demonstrate superior performance over baselines, while online AB tests at Xiaohongshu confirm significant improvements: +0.33% Time Spent, +8.18% CTR, +3.43% advertising value, and +1.19% GMV growth.
This work advances industrial recommendation systems by providing a scalable, efficient, and effective solution for multi-scenario ranking. It reduces engineering overhead, enhances information reuse, and supports diverse business objectives within a single unified framework. Future directions include model compression, multi-modal fusion, and adaptive scene modulation, aiming to further improve deployment efficiency and generalization across new scenarios. Overall, OneModel marks a significant step toward intelligent, resource-efficient, and user-centric large-scale recommender systems.
Deep Analysis
Background
随着互联网内容和电商平台的快速增长，推荐系统逐渐成为核心基础设施。早期多采用特征交互模型（如FM、DeepFM）进行单场景优化，随后引入序列建模（如DIN、Transformer）以捕获用户行为的长时序关系。近年来，生成式推荐模型（如HSTU、GenRank）利用长序列记忆和内容理解，显著提升效果。然而，这些模型多针对单一场景，难以满足多场景、多目标的工业需求，导致模型碎片化、资源浪费。多任务学习和多模型方案虽能缓解部分问题，但仍存在信息隔离、工程复杂等瓶颈。工业平台如小红书，用户在内容推荐、广告、商家服务等多个场景中行为连续，亟需一种统一架构，充分利用跨场景信息，提升效率与效果。
Core Problem
核心问题在于如何在保证模型性能的同时，支持多场景、多目标的工业部署。不同场景具有异构特征空间、目标冲突和干扰，单一模型难以兼顾专业化和迁移能力。传统多任务方案多采用浅层共享，难以捕获深层次关系，长序列建模成本高昂，部署复杂。此外，场景干扰导致迁移效果不佳，模型难以在实时系统中高效运行。解决这些问题需要创新架构，平衡信息共享与场景专业化，确保模型在复杂环境中稳定高效。
Innovation
本文创新点包括：1）场景感知信息调制（SAIM），通过门控机制减缓跨场景干扰，增强迁移能力；2）多目标端到端训练，结合自监督和监督目标，提升多场景适应性；3）分层用户表征，融合全局偏好与短期意图，增强个性化；4）采用Transformer的长序列解码器，支持大规模长序列建模，兼顾效率与效果；5）特征预取与图级推理，优化在线推理性能。这些创新结合，推动工业多场景推荐技术发展。
Methodology
- �� 行动导向的序列组织：将行为动作作为预测目标，物品作为上下文，减少序列长度。
- �� 场景感知特征投影：对不同场景的特征进行轻量级投影，映射到统一空间。
- �� 结构化上下文编码：引入场景、动作、时间、位置等信息，丰富行为表示。
- �� 长序列建模：采用Transformer解码器，结合因果掩码和候选掩码，支持多场景长序列学习。
- �� 场景感知调制：在FFN中引入门控机制，减缓跨场景干扰。
- �� 分层用户表征：融合全局偏好与局部意图，提升用户模型。
- �� 多目标训练：结合自监督和多任务目标，优化多场景性能。
- �� 线上推理优化：特征预取、用户状态缓存、图级推理，提升实时性。
Experiments
在小红书平台采集的真实数据集上，评估模型在多场景下的表现。指标包括AUC、LogLoss，比较单场景与多场景训练效果。设置不同上下文长度和模型容量，验证模型的扩展性。通过消融实验分析SAIM机制、分层表征和多目标策略的贡献。线上AB测试验证模型的实际效果与部署效率。结果显示，OneModel在多场景中优于传统模型，资源利用率显著提升。
Results
模型在离线指标上持续优于基线，广告CTR提升8.18%，GMV增长1.19%。AB测试中，探索流时间提升0.33%，互动率提升1.25%；广告CTR提升8.18%，广告价值增长3.43%；商家GMV提升1.1867%，GPM提升2.1585%。消融验证，SAIM和分层表征是性能提升关键。模型在不同上下文长度和容量下表现稳定，验证其实用性。
Applications
该模型适用于多场景推荐平台，支持内容、广告和商家服务的联合排序。依赖长序列行为数据，结合多目标优化，提升用户体验和商业转化。未来可扩展到多模态信息融合，支持更复杂场景，推动智能推荐发展。
Limitations & Outlook
模型训练成本高，长序列建模对计算资源要求大。在新场景或冷启动时表现仍有限，场景调制机制需优化。未来应研究模型压缩、推理加速和自适应调制策略，扩大应用范围。
Plain Language Accessible to non-experts
想象你在一个大型厨房里准备一顿丰盛的晚餐。每道菜代表一个不同的场景，比如主菜、甜点、饮料。厨师（模型）需要根据不同菜肴的特点，合理安排食材和烹饪时间。以前的厨师只专注于一道菜，效果有限。而这个新厨师（OneModel）可以同时处理多道菜，记住所有食材的搭配和时间，还能根据你的偏好（喜欢辣的还是甜的）调整调味料。它还能根据场合（家庭聚会或正式宴会）灵活应对。这样，不仅节省时间，也保证每道菜色香味俱佳。这个厨师就像模型一样，融合多场景信息，既专业又高效，能应对复杂厨房任务。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的游戏，里面有很多任务，比如打怪、收宝、交朋友。每个任务都需要不同策略，但有时它们会互相影响，比如交朋友帮你打怪，收宝帮你升级。以前的设计师会为每个任务写一个程序，但太麻烦，也难升级。现在，有个聪明的设计师（模型）发明了一套办法，让所有任务用一个大脑处理。这个大脑记住你所有的行动，知道你喜欢什么，在哪个任务表现最好，还能根据不同任务调整策略，比如打怪时变勇敢，交朋友时更善良。这样一来，无论你做什么，游戏都变得更聪明、更有趣，也更容易升级。这个大脑就像模型，能同时处理很多任务，还能学会你的偏好，帮你变得更厉害！
Abstract
Platform-scale recommender systems often span multiple business streams such as organic recommendation, advertising, and merchant services, where user behaviors form a continuous cross-stream trajectory. Maintaining separate ranking systems fragments user representations and increases engineering cost. We propose \textbf{OneModel}, a unified framework for multi-stream final ranking. OneModel maps heterogeneous behaviors into shared event sequences, learns long-context user representations with an action-oriented backbone, and introduces \emph{Scenario-aware Information Modulation} to balance cross-stream transfer and stream-specific specialization. For production deployment, OneModel further adopts stratified user representation, multi-objective training, and optimized online serving with feature decomposition, user feature prefetching, shared user-tower computation, and graph-level inference optimization. We deploy OneModel in production at \emph{Xiaohongshu}, where it delivers consistent offline gains over strong baselines and scales favorably with context length and model capacity. Online A/B tests improve Time Spent by \textbf{+0.33\%} and Engagement by \textbf{+1.25\%} in Explore Feed, lift advertising value by \textbf{+3.43\%} and CTR by \textbf{+8.18\%} in Feed Advertising, and raise DGMV by \textbf{+1.1867\%} and GPM by \textbf{+2.1585\%} in Merchant Recommendation, validating unified multi-stream ranking as an effective production foundation.