Anthropic Fellows Program — Reinforcement Learning

Details

Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. The Fellows Program provides funding and mentorship to promising technical talent to work on empirical AI research projects (with a public output such as a paper). Fellows primarily use external infrastructure (open-source models, public APIs) to complete a project aligned with Anthropic research priorities. This page refers specifically to the Reinforcement Learning workstream of the Anthropic Fellows Program (cohorts starting July 2026 and beyond).

Responsibilities

• Run an empirical research project using external infrastructure (e.g., open-source models, public APIs) with the goal of producing a public output (paper, report, code).
• Implement, evaluate, and iterate on RL environments and algorithms to improve model capabilities or safety-related behaviors.
• Build model-based tools to analyze and improve training data quality and study generalization.
• Collaborate with assigned mentors and the broader research and engineering teams; balance research exploration with engineering rigor and operational reliability.
• Use available compute funding and shared workspace resources; document and communicate research findings.

Requirements

• Fluent in Python programming (required).
• Available to work full-time on the Fellows program (expectation: 40 hours/week for 4 months).
• Work authorization and located in the United States, United Kingdom, or Canada for the program duration.
• Strong technical background in computer science, mathematics, or physics, or equivalent experience.
• Comfortable implementing ideas quickly and communicating clearly.

Strong candidates may also have

• Strong software engineering skills with experience building complex ML systems.
• Experience training, fine-tuning, or evaluating large language models.
• Experience with large-scale distributed systems and high-performance computing.
• Experience designing RL environments, RL algorithms, and analyzing/debugging model training processes.
• Background in related disciplines (e.g., economics, social sciences, cybersecurity) relevant to the workstream.

Mentors & Example Projects

Potential mentors for the Reinforcement Learning workstream include Ruhua Jiang, Kaidi Cao, Sunny Duan, David Brandfonbrener, Colt Steele, Dino Distefano, and Will Williams.

Example projects include building model-based tools for training-data analysis, research on generalization, creating RL environments to improve Claude models, building safety-focused RL environments, and implementing RL algorithm research.

Logistics

• Program duration: 4 months, full-time (with possible extension).
• Fellows must have work authorization in the US, UK, or Canada and be located in that country during the program.
• Shared workspaces are available in London and Berkeley; Anthropic is also open to remote fellows in the UK, US, or Canada. Applicants will be asked about availability to work from Berkeley or London (full- or part-time).
• Visa sponsorship is not available for the Fellows program; fellows must independently obtain or already have work authorization.

Compensation & Support

• Weekly stipend: 3,850 USD / 2,310 GBP / 4,300 CAD per week (expected base stipend).
• Expectation: 40 hours per week for 4 months.
• Funding for compute (~$15k/month) and other research expenses.
• Direct mentorship from Anthropic researchers, access to a shared workspace, connection to the AI safety and security research community, and other benefits that vary by country.

Interview & Application

• Application process includes initial application & reference check, technical assessments & interviews, and a research discussion.
• Apply via the Constellation application form linked on the posting. The next cohort starts July 20, 2026; apply by April 26, 2026 for that cohort (applications accepted on a rolling basis for later cohorts).