部门介绍
字节跳动搜索团队主要负责抖音、国际化短视频、今日头条、红果短剧、番茄小说、AI搜索等产品以及电商、生活服务等业务的搜索算法创新和架构研发工作。我们使用前沿的机器学习\大模型技术进行端到端建模并不断创新突破，同时专注于分布式系统、机器学习系统的构建和性能优化，从内存、Disk等优化到索引压缩、召回、排序等算法的探索，致力于支撑字节跳动持续多元化的产品创新和高速的业务增长、基于大模型革新和重塑下一代搜索技术体系，充分给同学们提供成长自我的机会。主要工作方向包括：
- 探索前沿的NLP技术：全面基于LLM的Query分析、相关性、个性化预估、满意度评估、生成式检索等，全链路应用LLM/VLM，每个细节都充满挑战；
- 跨模态匹配技术：进行多模态预训练，在超大规模的多模态表征、匹配、生成等多个技术方向上持续突破，打造世界领先的多模态搜索系统；
- 大规模流式机器学习技术：应用大规模机器学习，解决搜索中的推荐问题，让搜索更加个性化更加懂你；
- 千亿级数据规模的架构：从大规模离线计算，分布式系统的性能、调度优化，到构建高可用、高吞吐和低延迟的在线服务的方方面面都有深入研究和创新；
- 推荐技术：基于超大规模机器学习、因果推断、大模型推理等技术手段，构建业界领先的搜索推荐系统，对搜索推荐技术进行探索和创新。 课题介绍：随着大模型技术的快速发展，AI搜索领域迎来了新的机遇和挑战。传统搜索技术在面对海量数据、多模态信息以及用户多轮复杂需求时，开始暴露出很多问题。因此需要基于大模型来构建下一代AI搜索系统，提升搜索系统的智能化水平，优化用户体验，具体目标包括：
- 探索大模型与排序算法的结合，提升个性化排序的精度和用户体验；
- 探索基于多模态预训练的端到端生成式搜索大模型；
- 探索基于大模型Agent技术，提升复杂多义Query和多轮搜索下的用户满意度。 课题挑战
- 个性化排序的挑战：传统排序算法难以充分利用多模态信息，且模型复杂度有限，无法满足用户对精准化和个性化搜索的需求；
- 超大规模检索排序的挑战：传统的基于判别式的级联排序系统，难以满足搜索千亿级别候选的检索排序效率需求；
- 搜索需求日益复杂的挑战：用户搜索需求的复杂度在不断增加，传统搜索框架难以在多轮对话下，准确理解长难、多义Query的语义，导致搜索结果满意度低。 课题价值
- 技术价值：突破传统搜索技术瓶颈，构建大模型Agent驱动的下一代AI搜索架构，解决个性化排序、超大规模检索排序、复杂搜索需求理解满足等行业难题；
- 业务价值：大幅度提升搜索的用户体验和满意度，带动搜索场景LT和主动搜索心智的提升。 Topic Content With the rapid advancement of foundation model technology, the field of AI-powered search is encountering new opportunities and challenges. Traditional search technologies have begun to reveal significant limitations when confronted with massive data volumes, multimodal information, and complex multi-turn user needs. It is therefore necessary to leverage foundation models to build next-generation AI search systems, enhancing the intelligence of search systems and optimizing user experience. Specific objectives include:
- Explore the integration of foundation models with ranking algorithms to improve personalized ranking accuracy and user experience;
- Explore end-to-end generative search models based on multimodal pre-training;
- Explore LLM-based agent technology to improve user satisfaction under complex ambiguous queries and multi-turn search scenarios. Topic Challenges
- Personalized ranking: Traditional ranking algorithms struggle to fully leverage multimodal information, and their limited model complexity fails to meet user demands for precise and personalized search;
- Ultra-large-scale retrieval and ranking: Traditional discriminative cascaded ranking systems cannot meet the efficiency requirements for retrieval and ranking across hundred-billion-scale candidate pools;
- Increasingly complex search needs: User search needs are growing increasingly complex. Traditional search frameworks struggle to accurately understand the semantics of long, complex, and ambiguous queries in multi-turn conversations, resulting in low search result satisfaction. Topic Value
- Technical value: Break through the bottlenecks of traditional search technology; build a next-generation AI search architecture driven by LLM agents; and address industry challenges including personalized ranking, ultra-large-scale retrieval and ranking, and understanding and fulfilling complex search needs;
- Business value: Significantly improve search user experience and satisfaction, driving improvements in search LT and users' proactive search intent.
岗位职责
招聘官网暂未提供完整岗位职责，请前往官网核实。
职位要求
- 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；
- 具备优秀的编程能力，拥有扎实的机器学习/深度学习/大语言模型专业基础，熟练掌握C/C++或Python编程语言。 加分项：
- 具备出色的问题定义、分析与解决能力；在AAAI、NeurIPS、ACL等人工智能顶会有论文发表，或具备深厚科研经验者优先考虑；
- 有韧性，沟通能力与团队协作能力优秀；热爱技术，愿意与团队共同迎接挑战，具备创新钻研精神。
- Individuals who are completing or recently completed a PhD in AI, Computer Science, Computer Engineering, or a related technical discipline;
- Excellent coding skills and a solid foundation in Machine Learning/Deep Learning/LLM, proficient in C/C++ or Python. Preferred Qualifications
- Outstanding ability to define, analyze, and solve problems; candidates with publications at AI conferences such as AAAI, NeurIPS, ACL, etc., or in-depth research experience are preferred;
- Strong resilience, excellent communication and teamwork skills; passionate about technology, willing to embrace challenges with the team, and a drive for innovation.
