RL Scaling Research Scientist for Frontier Models

Periodic Labs; San Francisco (CA); On-site; USD 225,000 - 350,000; Full time

Job summary

Periodic Labs in Menlo Park, CA or Montreal, Canada is seeking a role focused on training frontier models with reinforcement learning to develop deep scientific knowledge and reasoning for scientific tasks. You will study how RL scales with training compute, model size, data, and reward quality, and work on algorithms for long-horizon RL.

The position requires hands-on RL experience, attention to detail, and the ability to transfer small-scale experiments to large-scale runs, within a complex

Qualifications

• Hands-on experience training LLMs with reinforcement learning.
• Strong attention to detail and rigorous approach to answer questions scientifically.
• Coming up with small-scale RL setups that transfer to large-scale training.
• Comfort working across a complex training stack to implement, debug, and test new research ideas.

Responsibilities

• Design experiments to understand RL performance scaling with compute, model size, data, and reward quality.
• Develop better RL algorithms across policy optimization, advantage estimation, exploration, and credit assignment for long-horizon tasks.
• Build adaptive sampling and curriculum methods adjusting task difficulty and rollout counts as models improve.
• Study bias and stability during RL training and address training–inference mismatch and policy staleness.
• Improve compute efficiency across training and inference through hyperparameter experiments.

Education

Bachelor's degree

Job description

Periodic Labs in Menlo Park, CA or Montreal, Canada is seeking a role focused on training frontier models with reinforcement learning to develop deep scientific knowledge and reasoning for scientific tasks. You will study how RL scales with training compute, model size, data, and reward quality, and work on algorithms for long-horizon RL.

The position requires hands-on RL experience, attention to detail, and the ability to transfer small-scale experiments to large-scale runs, within a complex