***Employment contracts can be signed in Hong Kong, but you must accept full-time or most of the time to work in Shenzhen!!!!
Duties:
Design or participate in the following research directions:
•Reasoning
•Computer Use Agent
•Code Agent
•Embodied Agent
1. Responsible for full-chain training in Code, Computer Use, Robotics scenarios, including but not limited to task construction, data collection, model training, evaluation, and improving model task execution performance
1. Reward Model optimization and innovation
2. Innovative exploration of new training paradigms such as r1-zero
3. Explore how to build a robust evaluation method, comprehensively, objectively and fairly evaluate the basic inference planning ability of the model, as well as its interaction ability with complex environments
2. Research data synthesis, scalable oversight, break through data bottlenecks, and reduce dependence on human annotation;
3. Study the application of system 2 in reasoning and planning abilities, use "slow thinking" to enhance effects, and optimize model basic abilities;
4. Enhance model's tool calling, API interaction ability, and solve complex problems by building agents.
Qualifications:
1. Artificial intelligence, computers, software engineering, electronic engineering, automation, robotics, mathematics, etc. Graduated with master's and doctoral degrees, those who are particularly outstanding may have their requirements relaxed.
2. Proficient in any one direction, such as computer vision, large language models, multi-modal large models, reinforcement learning, and agents.
3. Have experience in using PyTorch and other deep learning frameworks, familiar with distributed training frameworks (such as Megatron-LM and DeepSpeed), and have multi-machine multi-card distributed training experience.
4. Theoretical foundation is solid, with innovative spirit and in-depth thinking ability, strong communication skills and team collaboration spirit.
Have the following background preferred:
1. Have some accumulation in the following deep reinforcement learning fields
1. Model-Free Reinforcement Learning (Model-Free RL: Value-based Algorithm, Policy Gradients, Deterministic Policy Gradients, Distributional RL, Evolutionary Algorithms)
2. Imitation Learning/Inverse Reinforcement Learning (Imitation Learning and Inverse Reinforcement Learning: Behavior Clone, GAIL)
3. Exploration (Exploration: Intrinsic Motivation, Unsupervised RL)
4. Migration and multitask reinforcement learning (Transfer and Multitask RL: Progressive Networks, UVFA, UNREAL, HER)
5. Hierarchical Reinforcement Learning (Hierarchy RL: STRAW, Feudal Networks, HIRO)
2. Have knowledge and practical experience in realizing intelligent agents, combine long-term and short-term memory, retrieval augmented generation (RAG), and tool integration for dynamic environment.
3. Priority will be given to those with high-quality paper publications (such as ICML, NIPS, ICLR, ACL, CVPR, etc.), those with strong academic competition experience, those with significant influence in open-source communities, and those with engineering experience.
