SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Luma AI is seeking a Research Scientist / Engineer to build and scale distributed reinforcement learning post-training systems at frontier scale. You will own the full RL loop—training, rollout generation, environment execution, and reward computation—running concurrently across thousands of GPUs while maintaining speed, stability, and correctness.
Key Responsibilities:
- Design, build, and scale distributed RL post-training systems, orchestrating trainer, rollout, environment, and reward workloads across thousands of GPUs.
- Build high-throughput rollout generation, integrating inference engines (vLLM, SGLang), weight synchronization, and asynchronous/off-policy schemes.
- Design RL environments for agentic, multi-step tasks including sandboxed code execution, tool use, computer use, and multimodal interaction, reproducible and scalable to millions of episodes.
- Build reward infrastructure: verifiable/programmatic rewards, reward-model serving, LLM-as-judge pipelines, and defenses against reward hacking.
- Develop evaluation, monitoring, and debugging tooling that keeps large RL runs stable.
- Advance training efficiency and stability, and turn new post-training ideas into production runs with researchers.
First 90 Days:
Days 1–30: Immerse in and diagnose the current RL stack, identifying where throughput, stability, or correctness break. Days 30–60: Ship and validate improvements to a piece of the loop (rollout throughput, reward infrastructure, or an environment) on a real run. Days 60–90: Scale and systemize the full loop across thousands of GPUs and asynchronous architectures.
Requirements:
- Hands-on experience post-training LLMs with RL (PPO/GRPO-family, RLHF, RLVR) at meaningful scale.
- Extensive distributed PyTorch training and parallelism (FSDP, Tensor/Pipeline/Expert Parallel) for foundation models.
- Experience building RL environments, reward functions, verifiers, or evaluation harnesses for LLM agents, including sandboxed execution and multi-turn tool use.
- Deep familiarity with RL post-training frameworks (veRL, OpenRLHF, TRL, Ray orchestration) and rollout inference engines (vLLM, SGLang).
- Strong understanding of GPU clusters, networking, and communication libraries (NCCL, MPI) under mixed training and inference workloads.
Nice to Have:
- Running RL training across 100+ GPUs, including asynchronous or disaggregated trainer/rollout architectures.
- Containerization and orchestration (Kubernetes, Ray) for large environment fleets and sandboxed workloads.
- Research contributions in RL for LLMs, or open-source contributions to RL training frameworks.
