Research Engineer / Scientist - Reinforcement Learning

Skills

About the Role

You will develop and evaluate reinforcement learning algorithms and systems for reasoning, coding, and long-horizon tasks. Depending on your strengths, you will conduct algorithm research, study scaling behavior, or extend RL training frameworks. You will run controlled experiments, improve distributed and asynchronous training workflows, diagnose instability and regressions, and turn findings into robust training recipes.

Requirements

• Demonstrated depth in reinforcement learning algorithms, empirical research on model training, or distributed machine learning systems.
• Strong Python programming skills and experience with PyTorch or a comparable machine learning framework.
• Ability to design controlled experiments, interpret results, and debug unexpected model or system behavior.
• Experience taking a research idea or systems change through implementation, testing, and evaluation.
• Experience with distributed training frameworks or asynchronous rollout pipelines.
• Experience diagnosing unstable reinforcement learning runs, including policy-staleness or numerical issues.
• Experience working on coding agents, verifiable rewards, or long-horizon environments.

Responsibilities

• Develop and test reinforcement learning methods for reasoning, coding, and long-horizon tasks.
• Investigate stability in asynchronous reinforcement learning, including policy staleness, trainer-sampler mismatch, and numerical differences.
• Work with data and evaluation researchers to assess how task selection, rewards, and training data affect model learning.
• Design experiments on how reinforcement learning performance changes with model size, training compute, and task horizon.
• Use controlled ablations and evaluations to distinguish learning gains from reward exploitation, evaluation artifacts, and training bugs.
• Turn experimental findings into training recipes and validate them on larger models and longer runs.
• Extend prime-rl to support new algorithms, models, and asynchronous training workflows.
• Improve performance and numerical correctness across trainers, rollout workers, and environments.
• Build diagnostics for unstable runs and performance regressions.

Benefits

• Equity incentives.
• Flexible work arrangements, including remote work or in-person work in San Francisco.
• Visa sponsorship.
• Relocation assistance.
• Quarterly team off-sites.
• Hackathons.
• Conferences.