Research Engineer, RL Scaling Science

About the role

Anthropic's RL Scaling Science team studies how reinforcement learning behaves as we scale it (across model size, compute, and task horizon) and turns that understanding into the training recipes behind our frontier models. As a Research Engineer on this team, you'll design and run large-scale experiments to understand and resolve bottlenecks, build the benchmarks that make long-horizon progress measurable, and ship validated findings directly into production training.

This role lives at the boundary between research and engineering. The problems are open, the experiments run at frontier scale, and the path from a robust result to production is short.

Key responsibilities

• Design, run, and interpret large-scale RL experiments, reasoning rigorously about what the data does and doesn't show
• Investigate how RL improves as horizon, compute, and model size grow
• Build and maintain benchmarks for long-horizon RL so progress is measurable and reproducible
• Translate validated findings into production training recipes, exercising judgment about when a result is robust enough to ship
• Debug complex issues at the seam where research meets infrastructure - failures that only appear at scale
• Partner closely with adjacent RL teams across research and engineering and advance our overall RL stack

Minimum qualifications

• Strong empirical research skills in Reinforcement Learning, large-scale ML training, or a closely adjacent area
• Demonstrated ability to own large experiments end-to-end, from design through interpretation
• Proficiency in Python and experience working with large-scale or distributed ML systems
• Comfort operating at the research/systems boundary, including debugging where the two meet
• Care about the societal impacts of AI and responsible scaling

Preferred qualifications

• Published or shipped work in long-horizon RL or RL fundamentals
• Experience translating research findings into production training recipes
• Demonstrated large scale industry impact via RL interventions 
• Experience working on frontier-scale training runs with long trajectories

Representative projects

• Design a benchmark suite for long-horizon RL that distinguishes genuine capability gains from artifacts of evaluation setup
• Take a promising experimental finding, stress-test it across model scales, and work with training teams to land it in a production recipe
• Investigate an unexpected scaling trend in an RL run and trace it to a root cause spanning algorithm, data, and infrastructure