The posting
作为聚焦科学智能的前沿团队，我们正在组建一支极具野心的顶尖研究力量，致力于让大模型真正理解并解决最复杂的科学问题（如蛋白质折叠、量子模拟、多尺度物理仿真等）。 As a pioneering team focused on AI for Science (AI4S), we are assembling an ambitious, world-class research force dedicated to enabling Large Reasoning Models to truly understand and solve the most complex scientific challenges (e.g., protein folding, quantum simulation, multi-scale physical emulation).
在这里，你将参与构建具备多步逻辑推演、跨模态知识整合与自主假设验证能力的 深度推理（System 2 Reasoning）AI 系统，突破大模型在科学场景下的能力边界，共同定义科学发现的新范式。 In this role, you will help build System 2 Reasoning AI systems capable of multi-step logical deduction, cross-modal knowledge integration, and autonomous hypothesis verification. You will push the boundaries of LLMs in scientific contexts and help define a new paradigm for scientific discovery. 1、突破科学推理边界： 面向最具挑战性的科学问题（如分子动力学、蛋白质折叠、物理偏微分方程求解等），研发下一代大模型高效推理（LLM Reasoning）算法，提升模型在长逻辑链条下的科学推理能力。 Breaking Scientific Reasoning Frontiers: Develop next-generation LLM Reasoning algorithms for high-stakes scientific problems (Molecular Dynamics, Protein Folding, PDE solving) to enhance model performance across long logical chains. 2、攻坚后训练与对齐技术： 主导多模态大模型的 Post-training 流程，研究并改进指令微调（SFT）、强化学习（如 PPO, DPO等）算法，使模型能够精准对齐复杂的科学领域知识与人类专家意图。 Advancing Post-Training & Alignment: Lead the post-training pipeline for multimodal large models. Research and improve SFT and Reinforcement Learning (PPO, DPO, etc.) to precisely align models with complex scientific domain knowledge and expert intent. 3、构建科学数据与评测体系： 针对高难度科学问题，设计并构造高质量的训练/推理数据集，建立严谨、系统化的科学发现能力自动化评估与基准测试（Benchmark）体系。 Building Scientific Data & Evaluation Systems: Design and construct high-quality training/inference datasets for advanced science. Establish rigorous, systematic automated evaluation frameworks and benchmarks for scientific discovery capabilities. 4、全栈式研发与落地闭环： 提出具有高潜力的高原创性研究方向，完成从算法设计、原型验证、规模化分布式训练到系统部署的全流程。将技术落地于真实的生物医药、材料科学或物理仿真业务场景中，产生实际的科学与商业价值。 Full-Stack R&D & Deployment: Propose high-potential, original research directions. Manage the full lifecycle—from algorithm design and prototyping to large-scale distributed training and deployment—applying technology to real-world scenarios in biomedicine, materials science, or physics to generate tangible scientific and commercial value. 5、追求极致学术影响力： 将研究成果转化为行业标杆，在国际顶级期刊（如 Nature, Science）及人工智能顶级会议（如 NeurIPS, ICML, ICLR）上发表突破性论文，或开源具有全球影响力的模型与框架。 Pursuing Peak Academic Influence: Transform research into industry benchmarks.Publish breakthrough papers in top-tier journals (Nature, Science) and AI conferences (NeurIPS, ICML, ICLR), or release open-source models and frameworks with global impact.
1、具有计算机科学、机器学习、人工智能或相关领域的博士学位。 Ph.D. in Computer Science, Machine Learning, Artificial Intelligence, or a related field. 2、在大语言模型、科学智能、基座模型、强化学习、优化或统计等领域具备前沿研究经验，并作为第一作者在 NeurIPS、ICML、ICLR、AAAI 等人工智能顶会顶刊上发表过论文。 Frontier research experience in LLMs, AI for Science, Foundation Models, Reinforcement Learning, Optimization, or Statistics. Proven track record as a First Author in top-tier AI conferences or journals (NeurIPS, ICML, ICLR, AAAI, etc.). 2、深度掌握以下至少一项: Post-Training技术栈（SFT、DPO、PPO、GRPO），强化学习理论，多模态架构。 Deep mastery of at least one of the following: Post-Training stacks (SFT, DPO, PPO, GRPO), Reinforcement Learning theory, or Multimodal architectures. 3、具备扎实的软件工程能力，拥有大规模深度学习系统开发、GPU 集群算力优化与分布式训练实战经验。 Solid software engineering skills with hands-on experience in large-scale deep learning system development, GPU cluster optimization, and distributed training.
优先条件 Preferred Qualifications 1、跨学科复合背景： 对自然科学（生物、化学、物理、材料等）有强烈的探索欲，或曾参与过 AI 与基础科学交叉的深度研究项目。 Interdisciplinary Background: A strong curiosity for natural sciences (Biology, Chemistry, Physics, Materials) or prior experience in deep AI + Science cross-disciplinary projects. 2、Agentic System 经验： 在智能体（LLM Agents）规划、工具调用、反思与自我进化机制方面有深入研究或实际开发经验。 Agentic System Experience: Research or development experience in LLM Agents (planning, tool-use, reflection, and self-evolution mechanisms). 3、推理模型实战： 深入研究过大推理模型的推理架构，在 Test-time compute (TTC) 或强化学习探索机制上有独到见解。 Reasoning Model Expertise: Deep dive into reasoning architectures, with unique insights into Test-Time Compute (TTC) or RL exploration mechanisms. 4、拥有极强的好奇心、自我驱动力以及第一性原理思考能力，能够在高度不确定的前沿领域中寻找突破口。 Strong curiosity, self-drive, and First Principles thinking; the ability to find breakthroughs in highly uncertain, frontier domains.
