Research Scientist, Reinforcement Learning

Member of Technical Staff, Reinforcement Learning Research Our client is a well-funded, early-stage AI lab building a real-time, multimodal AI system designed to interact with the natural timing, emotion, and responsiveness of a human. The team is lean, highly credentialed, and moving fast on problems that are genuinely unsolved. About the Role Own RL and post-training for large-scale multimodal models at a frontier lab, from building the stack at 0→1 to scaling it to production. This is broader than a traditional RL algorithms role - you'll develop post-training methods and build the infrastructure needed to run them. The work spans rollout generation, reward modeling, policy optimization, evaluation, data feedback loops, serving, observability, and distributed execution. Responsibilities Build the RL/post-training stack from scratch: rollout generation, policy optimization, reward and reference model serving, data feedback loops, evaluation, checkpointing, and observability. Develop and scale post-training methods including PPO, GRPO, DPO, rejection sampling, RLHF/RLAIF, online RL, and model-based data improvement. Design systems abstractions connecting research ideas to production-scale RL runs: trainers, rollout workers, reward models, evaluators, data queues, experience buffers, and checkpoint promotion. Build evaluation and feedback loops for multimodal behavior: turn-taking, timing, emotional response, audiovisual coherence, instruction following, and real-time interaction quality. Optimize the end-to-end post-training loop across rollout throughput, serving latency, GPU utilization, policy update efficiency, and research iteration speed. Evolve the platform as algorithms, model architectures, reward definitions, data sources, and evaluation methods change. Qualifications PhD in ML, RL, CS, or a related field - completed or in final stretch Deep knowledge of RL and post-training methods: policy optimization, reward modeling, preference optimization, rejection sampling, KL control, and data feedback loops Ability to reason about training dynamics: reward hacking, unstable rewards, distribution shift, stale policies, mode collapse, over-optimization, noisy preferences, and evaluation mismatch Hands-on exposure to RL/post-training pipelines through research, internships, or open-source work; familiarity with frameworks such as verl, OpenRLHF, ms-swift, or equivalent, and rollout serving systems such as vLLM or SGLang Strong software engineering fundamentals with the appetite to build real systems, not just prototypes Must be willing to work on-site 5 days/week in Seattle, WA (relocation support available) Preferred Skills First-author publications at tier-1 ML journal/conference Internship experience at a top AI lab doing RL or post-training work Prior 0→1 experience building post-training systems, RL pipelines, agent training infrastructure, or evaluation platforms Hands-on experience with multimodal post-training for audio, video, or language models - particularly long-context or real-time interactive systems Experience with adjacent areas: distributed pretraining, inference serving, data infrastructure, simulation, or human/AI feedback collection Substantial open-source contributions in RL, post-training, alignment, or ML systems #J-18808-Ljbffr