小鹏汽车Research Scientist (World Model)

1、构建行业领先的世界模型，为具身场景提供生成式仿真能力。强化世界模型的长程时空记忆，物理属性模拟能力，实现可泛化、可落地、可scaling的世界模型，形成持续的技术影响力并引领国际行业发展。

We empower our people to stay resilient and relevant in a constantly changing world. We're looking for people who are always searching for creative ways to grow and learn. People who want to make a real impact, now and in the future. Does that sound like you? Then it seems like you'd make a great addition to our vibrant international team. DAI AIX – AI Acceleration and Exploration, is working on the cutting-edge research of Data Analytics and AI with Siemens global technology network, and consulting, co-creation, data driven applications for the end customers. Research Scientist is to do applied research for Industrial AI applications in the team. We are seeking a Reinforcement Learning (RL) Specialist to lead the design, implementation, and optimization of RL-driven systems for post-training of foundation models. The primary focus of this role is advancing our RL capabilities for real-world applications such as industrial control systems and LLM agents. You will develop cutting-edge algorithms, improve post-training efficiency, and deploy scalable RL solutions in industry. You'll make an impact by • 1. Reinforcement learning development for post-training: • Design and implement state-of-the-art RL algorithms (e.g., PPO, SAC, DQN) for post-training of foundation models like LLMs and time series foundation models. • Implement distributed RL training pipelines using frameworks like Ray RLlib, Deepspeed, or custom solutions. • Design and implement benchmark pipelines for model evaluation. • 2. Align foundation models like LLMs and time series foundation models with specific areas/tasks through techniques like SFT, RL. • 3. Coding & Infrastructure: • Write production-grade Python code using PyTorch, numpy, and pandas. • Manage Linux-based clusters for distributed training and deployment. • 4. All other support required by the line manager if necessary.

Model Optimization & Deployment: Design and implement efficient workflows for training, distillation, and fine-tuning Small and Large Language Models (SLMs), leveraging techniques such as LoRA, QLoRA, and instruction tuning. Apply model compression strategies—including quantization (e.g., GPTQ, AWQ) and pruning—to reduce inference costs and improve latency. Optimize LLM inference performance using frameworks like vLLM and TensorRT-LLM (TRT-LLM) to enable scalable, low-latency deployment. Build robust and scalable inference systems tailored to heterogeneous production environments, with a strong focus on performance, cost-efficiency, and stability. Evaluation & Data Management： Develop evaluation datasets and metrics to assess model performance in real-world product scenarios. Build and maintain end-to-end machine learning pipelines encompassing data preprocessing, training, validation, and deployment.  Cross-functional Collaboration： Collaborate closely with product managers, engineers, and research scientists to translate business needs into impactful AI solutions, driving real-world adoption and seamless product integration.

• Design and implement offline evaluation strategies that capture real-world usage and reflect end-user preferences. • Develop scientifically sound metrics that diagnose model regressions, benchmark against baselines (e.g., ChatGPT, Glean), and validate product improvements. • Manufacture synthetic yet realistic user activity data using LLMs to simulate diverse usage scenarios. • Collaborate on multi-agent systems or agentic workflows to automate evaluation flows and generate high-signal insights. • Analyze evaluation outputs to identify gaps in coverage, quality, and usability across Copilot canvases. • Partner with engineering and PMs to ensure insights are integrated into product workflows and experimentation pipelines. • Publish learnings in internal forums, external conferences, and contribute to best practices in applied science.