Reinforcement Learning Engineer

CoreWeave, the AI Hyperscaler™, acquired Weights Biases to create the most powerful end-to-end platform to develop, deploy, and iterate AI faster. Since 2017, CoreWeave has operated a growing footprint of data centers covering every region of the US and across Europe, and was ranked as one of the TIME100 most influential companies of 2024. By bringing together CoreWeave’s industry-leading cloud infrastructure with the best-in-class tools AI practitioners know and love from Weights Biases, we’re setting a new standard for how AI is built, trained, and scaled. The integration of our teams and technologies is accelerating our shared mission: to empower developers with the tools and infrastructure they need to push the boundaries of what AI can do. From experiment tracking and model optimization to high-performance training clusters, agent building, and inference at scale, we’re combining forces to serve the full AI lifecycle — all in one seamless platform. Weights Biases has long been trusted by over 1,500 organizations — including AstraZeneca, Canva, Cohere, OpenAI, Meta, Snowflake, Square,Toyota, and Wayve — to build better models, AI agents and applications. Now, as part of CoreWeave, that impact is amplified across a broader ecosystem of AI innovators, researchers, and enterprises. As we unite under one vision, we’re looking for bold thinkers and agile builders who are excited to shape the future of AI alongside us. If you're passionate about solving complex problems at the intersection of software, hardware, and AI, there's never been a more exciting time to join our team. Our Team The OpenPipe team at CoreWeave is building tools to help agents learn from experience . This is a critical step to make agents reliable enough to perform long tasks autonomously, in the same way human employees are. We’re systematically identifying and solving the major bottlenecks between today’s tech and those future self-improving agents. So far, we’ve: Released ART , the easiest library for getting started with RL. Developed RULER , a general-purpose reward function that works across many diverse tasks. Built Serverless RL , an elegant API that gives RL practitioners full control over their data, environment and reward function while letting them outsource the headaches of managing GPU infrastructure. These releases have a theme: we’re systematically tackling each major roadblock to successfully training self-improving agents. Several serious challenges remain. Building simulated environments often requires substantial human labor, and existing training methods are not data efficient enough. We're laser-focused on solving these problems and making self-improvement a reality for agent developers. In startup terms, this is a classic hard-tech bet. Our roadmap involves substantial technical risk ; there are still major technical problems we’re facing without a proven solution. However, there is very little market risk . We’ve worked closely with the teams building agents at many of the top AI-native startups as well as large enterprises. If we can build this, everyone will want it . A self-improving agent that learns from experience the way a human employee would could quickly capture a large fraction of the total inference market, which is worth tens of billions of dollars today and will be worth hundreds of billions in a few years. About the Role You have trained LLMs to be SOTA on specific tasks. You have opinions on whether sequence-level or token-level importance ratios are more effective. You probably shared the ScaleRL paper in your group chats, and kicked off a few ablations after you read it. This is an applied research role. You will be expected to generate and investigate research ideas towards solving the remaining obstacles to continuous learning in production. You will work with the broader OpenPipe team to validate these research directions across real customer tasks. We are very GPU rich and are ready to ... (truncated, view full listing at source)