岗位职责
- 基于VLM/Agent等 AI 能力，融合图文/视频笔记的文本、图像、视频、音频等多维度信号，覆盖笔记、评论、账号、交易、直播等多体裁场景的国际化内容风险识别与处置，持续提升主动召回能力，实现风险、体验、效率的协同优化；
- 建立并持续演进 Rednote 的内容特征体系与内容分级系统，输出结果应用于治理中的处置判罚以及搜推流控等核心链路，为 Rednote 全球生态的内容质量与安全构建长期竞争壁垒；
- 负责面向 Rednote 治理、内容理解、编辑等核心场景的 Agent 系统端到端研发，涵盖 ReAct / Tool Use / Multi-Agent 架构、意图理解与任务规划；深度参与 Agent 效果优化——从工具设计、上下文管理、编排策略到 Agentic RL 的 Post-Training，产出效果领先的 产品；
- 深度参与LLM/VLM 的预训练、SFT、RL 等全链路训练，参与构建多模态内容理解、编辑、生成等 AI 基础能力，推动 AI 从技术到业务价值的闭环落地；
- 持续追踪AI Agent领域的最新进展，引入并验证新技术的可行性，沉淀技术文档与最佳实践，推动前沿技术在rednote业务中的应用，保持团队的技术领先性，沉淀国际顶会论文（CVPR/NeurIPS/ICLR/ACL 等）。 【About the Team】 The International Trust & Safety (TnS) Algorithm Team is the safety foundation for rednote’s global expansion. We build content governance and understanding capabilities for a global community, spanning risk prevention, intelligent moderation, content understanding and creation. Built on our proprietary Zeus family of AI agents and VLM, we bring cutting-edge algorithms into real-world applications at scale, laying a strong foundation for rednote worldwide. Globalization opens new frontiers in community governance and content understanding. Cross-language and cross-cultural semantic understanding, the relationship between culturally diverse content creation and authentic expression, evolving regulations and compliance requirements across countries, and emerging risks in the AI era all call for new technical solutions. Here, we are building an AI-native content governance and understanding system for the world from the ground up, with our proprietary agents and large models powering the full lifecycle of risk detection, policy generation, and recursive self-improvement. Our goal is to build AI that understands every language and culture, empowering people to express themselves authentically and safely. If you are passionate about using cutting-edge algorithms to solve real-world challenges, join us in building a community where people around the world feel confident expressing themselves and can trust the space they share. 【Job Responsibilities】
- Multimodal Content Risk Detection and Mitigation: Leverage AI capabilities such as Vision-Language Models (VLMs) and AI Agents to integrate multimodal signals from text-and-image posts and videos, including text, images, video, and audio. Develop international content risk detection and mitigation solutions across diverse content and business scenarios, including posts, comments, accounts, transactions, and livestreams. Continuously improve proactive risk detection and recall while jointly optimizing risk control, user experience, and operational efficiency.
- Content Feature Engineering and Classification Systems: Build and continuously evolve rednote’s content feature systems and content classification frameworks. Deliver outputs that support key governance workflows, including enforcement decisions, as well as traffic control and distribution in search and recommendation systems. Establish a sustainable competitive advantage in content quality and safety across rednote’s global ecosystem.
- End-to-End AI Agent System Development: Lead and contribute to the end-to-end development of Agent systems for core rednote scenarios, including governance, content understanding, and content editing. Work across ReAct, tool use, and multi-agent architectures, as well as intent understanding and task planning. Drive end-to-end Agent performance optimization, covering tool design, context management, orchestration strategies, and Agentic RL post-training, to deliver state-of-the-art products.
- Full-Lifecycle LLM/VLM Training: Participate in the full model training lifecycle, including pre-training, supervised fine-tuning (SFT), and reinforcement learning (RL) for LLMs and VLMs. Help build foundational AI capabilities for multimodal content understanding, editing, and generation, and translate technical advances into measurable business value.
- Technical Exploration and Research: Stay up to date with the latest advances in AI Agent technologies, evaluate the feasibility of emerging approaches, and document technical findings and best practices. Drive the adoption of cutting-edge technologies across rednote’s business and contribute to research publications at leading international conferences, such as CVPR, NeurIPS, ICLR, and ACL.
职位要求
- 本科及以上学历，计算机科学与技术、人工智能、计算机、机器学习、大数据、数学等相关专业；
- 对大模型和Agent有较深入的理解，熟悉常见的LLM实施模式、PE、RAG、评估框架等，熟悉主流 Agent 架构（Multi-Agent、Context Engineering/Management、ReAct/PlanAct/CodeAct），了解Agent相关的交互协议（MCP、A2A、FunctionCall）；
- 熟悉Agentic RL等方法，了解相关 Infra（例如CodeRL，Reward System），能针对特定问题开展SFT、RL训练优化；
- 具备良好的学术视野和开拓精神，勇于钻研新技术，能主动思考、主动作为反馈。有CVPR/ICCV/NeurIPS/ICLR/ACL/EMNLP/KDD 等顶会论文发表经历更佳。
- Educational Background: Bachelor’s degree or above in Computer Science, Artificial Intelligence, Machine Learning, Big Data, Mathematics, or a related field.
- Strong Understanding of LLMs and AI Agents: In-depth understanding of large language models and AI Agents. Familiarity with common LLM implementation patterns, prompt engineering (PE), retrieval-augmented generation (RAG), and evaluation frameworks. Knowledge of mainstream Agent architectures and approaches, including multi-agent systems, context engineering and management, ReAct, PlanAct, and CodeAct. Familiarity with Agent interaction protocols and interfaces, including MCP, A2A, and function calling.
- Experience with Agentic Reinforcement Learning: Familiarity with Agentic RL and related infrastructure, such as CodeRL and reward systems. Ability to apply supervised fine-tuning (SFT) and reinforcement learning (RL) to optimize models for specific problems.
- Research Acumen and Innovation: Strong research awareness, intellectual curiosity, and a willingness to explore emerging technologies. A proactive mindset, with the ability to think independently, take initiative, and provide constructive feedback. Publications at leading conferences such as CVPR, ICCV, NeurIPS, ICLR, ACL, EMNLP, or KDD are a strong plus.
