Research Engineer / Research Scientist, RL Frontiers

Anthropic•San Francisco, NY

17h•Hybrid

About The Position

Reinforcement learning is how Claude learns to reason, write code, and act autonomously over long horizons. The RL Scaling team works on how RL scales: what happens to throughput, stability, and learning efficiency as models get larger, episodes get longer, and compute grows by orders of magnitude, and what has to change in our algorithms and systems to keep getting returns from that scale. This role sits squarely across research and engineering. You'll develop next-generation architectures and RL algorithms, take them from a small-scale result to a frontier-scale run, and understand every place they behave differently along the way. You'll build the systems that set how fast the team can iterate: how many experiments, at what scale, and how quickly we can trust the results. And you'll work on Anthropic's largest and fastest RL runs, where the gap between a good idea and a working one is often a problem no one has solved yet.

Requirements

Deep familiarity with modern transformer language models, including their architecture, training dynamics, and the behavior of large-scale optimization
Hands-on experience training large models in a distributed setting, including the tradeoffs between data, tensor, and pipeline parallelism
A track record of original technical work in ML training or systems, such as new methods, architectures, or optimizations, demonstrated through research, open-source, or production impact
Ability to design rigorous experiments at scale, including baselines, ablations, and enough statistical care to trust a result that costs real compute
Ability to reason quantitatively about the compute, memory, and communication costs of a model or algorithm
Strong programming skills in Python and JAX or PyTorch, and comfort reading and changing code at every layer of the stack

Nice To Haves

Research experience in reinforcement learning, optimization, or large-scale training, published or otherwise
Experience developing RL algorithms for language models
Experience with scaling laws or other quantitative models of training efficiency
Experience designing or modifying transformer architectures beyond standard configurations
Experience scaling training to large fleets of accelerators and debugging the problems that only appear at scale
Deep understanding of numerics in large-scale training, including low-precision formats and sources of instability
Familiarity with how GPU or TPU performance characteristics shape architecture and algorithm choices
Experience with C++ or Rust

Responsibilities

Study how RL training and sampling scale with model size, context length, and compute, and find the algorithmic and systems changes that keep scaling efficient
Develop next-generation model architectures and RL algorithms, and make them run efficiently at frontier scale
Take promising small-scale results to frontier-scale runs, and diagnose why they behave differently when they get there, whether the cause is numerical, algorithmic, or systemic
Build the experimental infrastructure that sets research velocity: fast, reproducible comparisons of architecture and algorithm variants at meaningful scale
Own end-to-end performance of our largest RL runs, from research code down to the hardware
Build performance and cost models for proposed architecture and algorithm changes, and use them to decide which ideas get scaled
Investigate training dynamics at scale, including instabilities, divergence, and throughput regressions, and trace them to root cause