The posting
团队介绍：推荐架构团队支撑字节跳动旗下多款APP产品，如抖音、今日头条、番茄小说、西瓜视频、剪映等推荐系统架构的设计和开发，保障系统的稳定和高可用，致力于抽象系统通用组件和服务，建设推荐中台、数据中台；关于在线服务，在这里你有机会参与大规模机器学习在线预估框架的研发与优化，也有机会参与模型训练与调度等相关问题的研究与突破，解决系统瓶颈，降低成本开销；如你对大数据感兴趣，在这里也有机会参与通用实时计算系统的开发、构建统一的推荐特征中台，为推荐业务实现先进的消重、计数、特征服务等；我们期待热爱技术的你加入，一起创造更多可能。
课题介绍：随着业务场景日趋复杂，推荐/广告/搜索面临严峻挑战。大模型虽能精准捕捉用户偏好、提升个性化效果及内容质量，但实时性、稳定性和可扩展性要求极高，在分布式训练、推理加速、异构硬件及多模态数据处理等方面存在诸多难点。同时，模型规模扩大与多模态数据涌现，使现有基础设施在数据处理效率与资源利用上难以满足需求。 本课题聚焦系统与工程创新，突破关键技术瓶颈，构建高效、稳定、可扩展的大模型解决方案，为搜广推场景提供坚实技术支撑。
课题挑战 1、大模型原生训推架构重构； 2、极致性能优化与AI Infra创新； 3、端到端生成式范式创新； 4、多模态AU数据基础设施与质量优化； 5、多模态数据表征和 RAG 应用系统。
课题价值 为搜广推业务构建下一代生成式AI基础设施，通过大模型、多模态技术与系统协同创新，突破性能瓶颈，实现超长上下文、毫秒级响应和极致信息精度，驱动业务智能升级。
Topic Content As business scenarios become increasingly complex, search, advertising, and recommendation are facing significant challenges. While large models can accurately capture user preferences and enhance personalization as well as content quality, they also impose stringent requirements on real-time performance, stability, and scalability. This introduces substantial technical challenges in areas such as distributed training, inference acceleration, heterogeneous hardware utilization, and multimodal data processing. At the same time, the rapid growth in model scale and the proliferation of multimodal data have made it difficult for existing infrastructure to meet the demands of data processing efficiency and resource utilization. This topic focuses on system and engineering innovations to overcome key technical bottlenecks and build efficient, stable, and scalable large-model solutions, providing a robust technical foundation for search, advertising, and recommendation scenarios.
Topic Challenges 1.Native training and inference architecture redesign for LLMs; 2.Extreme performance optimization and AI infrastructure innovation; 3.End-to-End generative paradigm innovation; 4.Multimodal AU data infrastructure and quality optimization; 5.Multimodal data representation and RAG-based application system.
Topic Value Building next-generation generative AI infrastructure for search, advertising, and recommendation businesses. Through the co-design of large models, multimodal technologies, and system-level innovations, we aim to overcome performance bottlenecks and enable ultra-long context handling, millisecond-level response latency, and high-precision information understanding, thereby driving intelligent upgrades across the business.
1、2027届毕业，获得博士学位，人工智能、计算机科学、计算机工程及相关技术专业； 2、具备优秀的编程能力，熟练掌握数据结构与基础算法；至少掌握C/C++/Python中一种语言，同时具备扎实的工程思维，能够兼顾性能与成本的平衡。
加分项： 1、在VLDB、SIGIR、OSDI等CCF A类期刊或会议有论文发表者优先； 2、有韧性，具备良好的沟通能力与团队协作能力；热爱技术，愿意与团队共同迎接挑战，富有创新精神。
1.Individuals who are completing or recently completed a PhD in Artificial Intelligence, Computer Science, Computer Engineering, or a related technical discipline; 2.Excellent programming abilities with a strong command of data structures and fundamental algorithms; For traditional coding roles, proficiency in C/C++ is required; for intelligent coding roles, proficiency in Python is required; Candidates are required to use these languages to implement complex algorithms and build iterative models; Candidates should also have a strong engineering mindset with the ability to balance performance and cost.
Preferred Qualifications 1.Candidates with publications in CCF-A category journals or conferences such as VLDB, SIGIR, OSDI, etc., are preferred; 2.Strong resilience, excellent communication and teamwork skills; passionate about technology, willing to embrace challenges with the team, and a drive for innovation.
