AI Agent Model & Algorithm Engineer / Specialist - Accio

Job Description:Accio is a strategic, AI-native application product incubated within Alibaba International Digital Commerce Group (AIDC / Alibaba.com). As the world’s first B2B AI Agent, Accio achieves leapfrog development by continuously exploring cutting-edge technologies such as Agents, LLMs, VLMs, RL, Memory, Reasoning, and AI Search to develop in-house Agent models and Agent systems.Responsibilities:Lead the design of Agent algorithmic architectures and optimize performance across domains including model optimization (Agentic Post-training, Agentic Infra, etc.), autonomous planning, multi-step reasoning, tool/skill usage, long/short-term memory, and RAG enhancement.Design and implement highly scalable Multi-Agent collaboration frameworks (e.g., Agent Swarm / Agent Team) for dynamic orchestration of complex tasks, enabling recursive decomposition of vague macro objectives into atomic tasks.Design long-horizon execution mechanisms by building multi-tiered storage systems combining working memory and global knowledge bases to solve context degradation in long-running Agents; establish shared cross-Agent context capabilities and design self-evolving Agent frameworks.Build end-to-end Agent evaluation systems and develop productivity-focused benchmarks to drive the high-impact deployment of Agents in real-world business scenarios.Explore and industrialize frontier Agent technologies, including but not limited to Agentic Models, Agentic Benchmarks, Agentic RL, Pro-active Agents, Function Calling, Tool-Use, Multi-Step Reasoning, Agent Harness, and Agentic Post-Training.Push the boundaries of Agent Architectures/Structures to maximize model capabilities within applied Agent research, exploring Self-Evolving AI Systems to achieve Self-Improving Agents.职位描述：Accio是阿里巴巴国际数字商业集团阿里国际站内部孵化的一款战略级AI原生应用产品，也是全球首个B2B AI Agent，通过持续探索Agent、LLM、VLM、RL、Memory、Reasoning、AI Search等前沿技术，自研Agent模型、Agent系统，实现B2B AI Agent跨越式发展。1、负责Agent算法架构设计与效果优化，包括但不限于模型优化（Agentic Post-training/Agentic Infra等）自主规划（Planning）、多步推理（Reasoning）、工具/skill调用（Tool Use）、长短期记忆（Memory）及 RAG 增强；2、复杂任务动态编排：设计并实现高扩展性的Multi-Agent协作框架（如Agent Swarm/Agent Team），支持将模糊的宏观目标递归拆解为原子任务；3、Long-horizon运行设计：构建结合"工作记忆+全局知识库"的多级存储系统，解决Agent长期运行中的信息遗忘问题，建立跨Agent的共享上下文能力，设计Agent持续进化框架；4、构建端到端的Agent评测体系，构建生产力场景benchmark，推动Agent在business场景落地；5、探索落地前沿Agent技术，包含而不限于：Agentic Model、Agentic Benchmark、Agentic RL、Pro-active Agent、Function Calling、Tool-Use、Multi-Step Reasoning、Agent Harness、Agentic Post-Training；6、探索Agent Architectures/Structures的上限，在Agent应用研究中最大程度释放模型的能力，研究Self-Evolving AI System，实现Self-Improving Agents。Job Requirements:Master’s degree or above, preferably in Computer Science, Electronics, Artificial Intelligence, Automation, Mathematics, Physics, or related disciplines.Solid engineering capabilities, with proficiency in development frameworks using Python or related technologies.Deep understanding of LLMs and AI Agents. Candidates with experience in post-training, execution paradigms, prompt engineering, evaluation frameworks, or retrieval architecture are preferred. Preference will be given to candidates with experience in:Agent data synthesis, Agentic RL training, and Agentic RL Infra (e.g., CodeRL, Reward Systems).Building Coding/Search Agents, Agent memory systems, or multi-Agent collaboration architectures.Agent Architectures/Structures (e.g., Multi-Agent systems, Context Engineering/Management, ReAct/PlanAct/CodeAct).Agent interaction protocols (e.g., MCP, A2A, Function Calling).Strong self-drive, quick learning capability, innovative mindset, and resilience under pressure, with the agility to adapt in the fast-evolving AI era.Preferred Qualifications / Bonus Points:Solid foundation in machine learning, with familiarity in CV, NLP, RL, ML, multimodal understanding, or Search/Recommendation/Ads; publications in top-tier conferences/journals (e.g., ICML, ACL, EMNLP, CVPR, ECCV, ICCV, NeurIPS, ICLR, KDD, WWW, SIGIR, RecSys, SIGGRAPH, or SIGGRAPH Asia).Work or internship experience in high-impact AI/Agent projects.Proven track record of contributing to influential open-source projects on platforms like GitHub.Experience developing or optimizing core algorithms for SOTA Agent products or deploying foundational LLM applications in production.Exceptional coding proficiency (C/C++ or Python); award winners in algorithmic competitions (e.g., ACM/ICPC, NOI/IOI, TopCoder, Kaggle).Proven track record of leading high-impact projects in LLMs, Multimodality, RL, Foundation Models, or World Models.职位要求：1、硕士及以上学历，计算机、电子、人工智能、自动化、数学、物理、等相关专业优先；2、具有良好的工程能力，熟悉Python等开发框架；3、对大模型和Agent有较深入的理解，熟悉常见的大模型后训练、实施模式、提示工程、评估框架、检索框架者优先。包含而不限于具有Agent数据合成、Agentic RL训练、Agentic RL Infra（例如CodeRL，Reward System）、Coding/Search Agent构建、Agent记忆系统构建、Multi-agent协作设计者优先；了解Agent Architectures/Structures（Multi-Agent、Context Engineering/Mangemant、ReAct/PlanAct/CodeAct）者优先；了解Agent相关的交互协议（MCP、A2A、Function Calling）者有优先；4、有较强的自驱力、学习力、创新力和抗压能力，能够跟上正在快速变化的AI时代。加分项1、具有扎实的机器学习基础，熟悉CV、NLP、RL、ML、多模态理解、搜推广等领域的技术，在ICML、ACL、EMNLP、CVPR、ECCV、ICCV、NeurIPS、ICLR、KDD、WWW、SIGIR、Recsys、SIGGRAPH或SIGGRAPH Asia等顶级会议/期刊上发表论文者优先；2、在具有影响力AI/Agent项目实习/工作经验者优先；3、在GitHub等开源社区具有影响力项目发表经验者优先；4、具有SOTA Agent产品核心算法开发/优化、大模型基座应用落地经验者优先；5、具有优秀的代码能力，熟练掌握C/C++或Python，在ACM/ICPC、NOI/IOl、Top Coder、Kaggle等比赛获奖者优先；6、在LLM、多模态、RL、基础模型、世界模型等领域，主导过大影响力项目者优先。