This is a fully remote position, open to applicants in Europe.

📋 Description

• Design, develop, and enhance distributed reinforcement learning post-training systems utilizing thousands of GPUs.

• Create high-throughput rollout generation systems that incorporate vLLM, SGLang, weight synchronization, and asynchronous/off-policy strategies.

• Engineer scalable reinforcement learning environments for agentic, multi-step tasks, including sandboxed code execution, tool usage, computer interaction, and multimodal interfaces.

• Construct reward infrastructure featuring verifiable/programmatic rewards, reward-model serving, LLM-as-judge pipelines, and safeguards against reward manipulation.

• Create evaluation, monitoring, and debugging tools to ensure stable large-scale reinforcement learning operations.

• Enhance training efficiency and reliability while transforming post-training concepts into production implementations alongside researchers.

• Acquire knowledge of the existing RL stack, identify bottlenecks, implement and validate enhancements, and reinforce the complete loop across thousands of GPUs.

⛳️ Requirements

• Practical experience with post-training language models leveraging reinforcement learning (PPO/GRPO-family, RLHF, RLVR) at significant scale.

• In-depth experience with distributed PyTorch training and parallelism (FSDP, Tensor/Pipeline/Expert Parallel) aimed at foundation models.

• Proven track record in constructing reinforcement learning environments, reward functions, verifiers, or evaluation harnesses for language model agents, including sandboxed execution and multi-turn tool use.

• Profound understanding of RL post-training frameworks (veRL, OpenRLHF, TRL, Ray orchestration) and rollout inference engines (vLLM, SGLang).

• Strong grasp of GPU clusters, networking, and communication libraries (NCCL, MPI) under mixed training and inference scenarios.

• Experience with containerization and orchestration (Kubernetes, Ray) for large fleets of environments and sandboxed workloads.

• Contributions to research in reinforcement learning for language models, or open-source contributions to RL training frameworks.