Multimodal Reinforcement Learning Post-Training Algorithm Expert

This role focuses on multimodal reinforcement learning post-training algorithms. The expert will bridge algorithm and framework teams, translating advanced post-training principles (RLHF, DPO) for large models into framework requirements. Responsibilities include optimizing and evaluating post-training pipelines for stability, efficiency, and generalization, with a focus on cross-modal alignment and reward function design. The role also involves researching cutting-edge advancements, troubleshooting training bottlenecks, and collaborating on solutions. Efficient cross-team support, technical documentation, and knowledge sharing are essential to enhance team expertise.

Must Have

• Master's degree or higher in Computer Science, Artificial Intelligence, Electronic Engineering, Automation, or related fields.
• Solid foundation in machine learning/deep learning.
• Deep understanding of multimodal large models and reinforcement learning post-training technology stack.
• Proficiency in Python programming and familiarity with deep learning frameworks like PyTorch.
• Deep understanding of model architectures such as Transformer and Diffusion.
• Thorough comprehension of SFT, RLHF, and DPO post-training algorithms.
• Strong engineering implementation and debugging skills.
• Capable of rapidly validating algorithmic ideas and conducting rigorous experimental analysis.
• Familiarity with at least one mainstream large model training/inference framework (e.g., Megatron-LM, DeepSpeed, VLLM).
• Excellent cross-team communication skills.
• Strong sense of responsibility, self-motivation, and passion for solving complex problems.

Good to Have

• Experience with post-training frameworks like VERL or OpenRLHF is a plus.

Business Unit

Technology Engineering Group (TEG) is responsible for supporting the company and its business groups on technology and operational platforms, as well as the construction and operation of R&D management and data centers, TEG provides users with a full range of customer services. As the operator of the largest networking, devices, and data center in Asia,TEG also leads the Tencent Technology Committee in strengthening infrastructure R&D through internal and distributed open source collaboration, constructing new platforms and supporting business innovation.

What the Role Entails

• Algorithm-Framework Co-design: Act as a technical bridge between the algorithm and framework teams. Deeply understand the principles and evolution trends of post-training algorithms for multimodal large models (e.g., RLHF, DPO, Curriculum Reinforcement Learning) and translate these into functional requirements for the underlying frameworks, providing insights for framework architecture design
• Training Pipeline Optimization and Evaluation: Lead or deeply participate in the setup, optimization, and effectiveness evaluation of post-training pipelines (e.g., multimodal SFT, RLHF). Focus on training stability, efficiency, and generalization capability, particularly proposing systematic improvements for areas like cross-modal alignment, reward function design, and policy optimization
• Technical Research and Bottleneck Resolution: Proactively track cutting-edge advancements in multimodal reinforcement learning post-training from academia and industry. Perform root cause analysis for training bottlenecks (e.g., insufficient OOD generalization, modality fusion conflicts) and collaborate with the framework team to develop and implement solutions
• Cross-team Support and Knowledge Sharing: Collaborate efficiently with framework development, hardware optimization, and business algorithm teams to ensure the implementation of technical solutions. Produce high-quality technical documentation, design drafts, and experimental reports. Organize internal sharing sessions to enhance the overall technical expertise of the team

Who We Look For

• Education and Technical Background: A Master's degree or higher in Computer Science, Artificial Intelligence, Electronic Engineering, Automation, or related fields. A solid foundation in machine learning/deep learning, with a deep understanding of multimodal large models and the reinforcement learning post-training technology stack

Core Algorithm and Engineering Skills:

• Proficiency in Python programming and familiarity with deep learning frameworks like PyTorch.
• Deep understanding of model architectures such as Transformer and Diffusion
• Thorough comprehension of the principles, processes, and common challenges (e.g., training instability, reward hacking) of post-training algorithms like SFT, RLHF, and DPO
• Strong engineering implementation and debugging skills, capable of rapidly validating algorithmic ideas and conducting rigorous experimental analysis for performance evaluation

Framework Collaboration and System Perspective:

• Familiarity with at least one mainstream large model training/inference framework (e.g., Megatron-LM, DeepSpeed, VLLM) and an understanding of their architectural design principles
• Ability to assess framework usability, scalability, and performance from an algorithmic perspective and propose improvement suggestions. Experience with post-training frameworks like VERL or OpenRLHF is a plus
• Soft Skills: Excellent cross-team communication skills, able to clearly translate requirements and articulate solutions between algorithm and engineering teams. A strong sense of responsibility, self-motivation, and passion for solving complex problems