Research Internship Reinforcement Learning (Summer)

Who are we? Cohere is the leading security-first enterprise AI company. We build cutting-edge foundation AI models and end-to-end products that are designed to solve real-world business problems. We’re training and deploying frontier models for enterprises who are building AI systems. We believe that our work is instrumental to the widespread adoption of AI and we are looking for folks that want to be part of that. We obsess over what we build. Each one of us is responsible for contributing to increasing the capabilities of our models and the value they drive for our customers. Cohere is a team of researchers, engineers, designers, and more, who are all passionate about their craft. We are a global technology company co-headquartered in Toronto and San Francisco, with key offices in London, New York City, Montreal, Seoul, Germany and Paris. Join us! Duration: Minimum 4 months (summer 2026, with potential extension) ABOUT THE PROJECT This internship offers a unique opportunity to contribute to cutting-edge research in reinforcement learning (RL) and large language models (LLMs), focusing on two interconnected projects: 1. Combining Self-Distillation and Reinforcement Learning for LLMs, with Applications to Code and Agentic Tasks This project explores how LLMs can improve through self-reflection and iterative learning by combining reinforcement learning with verifiable rewards (RLVR) and self-distillation. The focus is on scenarios where structured feedback from verifiers, compilers, unit tests, or tool calls enables models to detect errors, revise outputs, and learn from failures. The internship will bridge theoretical mathematical modeling of self-distillation with practical, production-oriented implementation. 2. Dealing with Extremely Large Rollouts in RLVR As RLVR becomes a cornerstone for training reasoning-oriented LLMs, the challenge of handling extremely large rollouts grows. This project investigates mechanisms such as summarization, memory, context compaction, hierarchical sub-agents, and resumable rollouts to enable unbounded or very long trajectories. It also explores how to effectively learn from such trajectories, as traditional RLVR objectives fail when episodes exceed context window limits. Both projects are grounded in recent research and aim to advance the state-of-the-art in LLM training and deployment. Responsibilities - Conduct literature reviews and implement state-of-the-art algorithms in RL and self-distillation. - Design and execute experiments to evaluate the effectiveness of proposed methods on code generation and agentic tasks. - Develop and maintain codebases for both theoretical modeling and practical implementations. - Collaborate with researchers to analyze results, refine methodologies, and prepare findings for publication. - Contribute to the design of mechanisms for handling large rollouts, such as summarization and hierarchical sub-agents. - Document progress, methodologies, and outcomes clearly and comprehensively. REQUIREMENTS - Technical Skills: - Strong background in machine learning, particularly reinforcement learning and deep learning. - Proficiency in Python and experience with ML frameworks (e.g., PyTorch, TensorFlow). - Familiarity with LLMs and their training paradigms. - Experience with coding tasks, unit testing, or compiler tools is a plus. - Educational Background: - Currently pursuing a Master’s or PhD in Computer Science, Machine Learning, or a related field. - Soft Skills: - Ability to work independently and manage complex projects. - Strong problem-solving and analytical skills. - Excellent communication skills for collaborating with a research team. - Additional: - Prior experience with RLVR, self-distillation, or large-scale ML experiments is highly desirable. - Willingness to learn and adapt to new methodologies and tools. HOW AND WHERE WE WORK: - Cohere is remote-friendly. We have offices in Toronto, San Francisco, New York City, London, Paris, Montreal, and more coming