Research Engineer / Research Scientist, RL Frontiers
About this role
To keep reinforcement learning efficient as model sizes, context lengths, and compute requirements grow by orders of magnitude, engineers must overcome complex bottlenecks in throughput, stability, and algorithmic scaling. Anthropic’s RL Scaling team addresses these challenges by investigating how training dynamics change at frontier scale and designing the next generation of architectures and algorithms that empower Claude to reason, write code, and act autonomously.
Bridging the gap between empirical research and robust systems engineering, team members take promising small-scale results and scale them up while diagnosing numerical, algorithmic, and systemic discrepancies along the way. Individuals in this function build experimental infrastructure, establish performance and cost models, and oversee end-to-end execution for the largest reinforcement learning runs. Successful contributors typically possess a strong quantitative background, deep expertise in distributed training and modern transformer architectures, and the ability to design rigorous, large-scale experiments.
What you will do
- Investigate how reinforcement learning training and sampling scale relative to model size, context length, and compute resources.
- Design and implement next-generation model architectures and reinforcement learning algorithms optimized for frontier-scale execution.
- Scale promising small-scale results to massive runs while diagnosing numerical, algorithmic, or systemic behavioral differences.
- Build experimental infrastructure to facilitate fast and reproducible comparisons of architecture and algorithm variants.
- Manage the end-to-end performance of large-scale reinforcement learning runs from research code down to underlying hardware.
- Develop performance and cost models to evaluate proposed architecture and algorithm changes before scaling.
- Trace training instabilities, divergence, and throughput regressions to their root causes across large fleets of accelerators.
What you need
- Demonstrate deep familiarity with modern transformer language models, including architecture, training dynamics, and large-scale optimization.
- Possess hands-on experience training large models in a distributed setting across data, tensor, and pipeline parallelism.
- Show a proven track record of original technical work in machine learning training or systems through research, open-source contributions, or production impact.
- Exhibit the ability to design rigorous experiments at scale using proper baselines, ablations, and statistical care.
- Reason quantitatively about the compute, memory, and communication costs associated with models and algorithms.
- Bring strong programming skills in Python alongside proficiency in JAX or PyTorch with comfort navigating every layer of the stack.
Nice to have
- Research experience in reinforcement learning, optimization, or large-scale training
- Experience developing reinforcement learning algorithms specifically for language models
- Familiarity with scaling laws or quantitative models of training efficiency
- Experience designing or modifying transformer architectures beyond standard configurations
- Deep understanding of numerics in large-scale training, including low-precision formats
Skills and tech stack
Questions candidates ask
What is the remote work policy?
Anthropic maintains a location-based hybrid policy where staff are expected to be in one of the offices at least 25% of the time, though some roles may require more in-office time.
What is the compensation range for this role?
The annual base salary range for this position is $500,000 to $850,000 USD.
Does the company provide visa sponsorship?
Anthropic does sponsor visas and retains an immigration lawyer to help, making every reasonable effort to secure a visa if an offer is extended.
What kind of team environment should I expect?
The organization operates as a single cohesive team working on big science research efforts, valuing open collaboration, frequent research discussions, and strong communication skills.