Research Scientist

We\'re building the autonomous training gym for RL agents. Bring an agent and a goal — the gym assembles the environment, trains the policy, and proves it works. For large models, LLM agents, robots, games, and science. The shared public corpus is spent. Every frontier lab is forking into proprietary experience: Anthropic is discussing $1B+/yr on RL environments, Moonshot trained Kimi K3 across 20,000+ tool environments, Harvey hand-built a 1,200-task legal environment with 75,000 expert rubrics just to train its own agents. Training is no longer the bottleneck — **environments are** . Our bet: **RL can teach an agent anything that can be practiced and scored.** Open hubs offer thousands of generic environments — but the environment that matters is yours: your tools, your data, your criteria of success. That one is still built manually, by every team. So we\'re building **arkenOS** , the operating system where environments, experiments, and results live — scaling to thousands of parallel GPU rollouts — and **Optimus** , the AI engineer that runs inside it: it assembles environments, writes rewards and tests, runs experiments, and fixes what fails. A goal goes in; a verified policy comes out. We\'re founded by the contributors to Newton and MuJoCo, with a founding team joining from Hippocratic AI, Harper Insurance (YC), and Meta SuperIntelligence Lab. This is the ground floor of the machine that builds the machines. ## About the role You own the research core of the autonomous loop. The open questions we\'re paid to answer are the hardest and most interesting ones in RL right now: how does an agent author a reward that can\'t be hacked? How does it know a training run is failing before the curves say so? What does "proven" mean, rigorously enough that a customer ships the policy? What you\'ll do ## What you\'ll do Drive research on agentic RL: online RL for LLM agents, reward modeling, RL from human demonstrations, and automated curriculum/environment generation. Build Optimus\'s core competencies — reward authoring, reward-hacking detection, failure diagnosis, and automated repair of training runs. Design the verification science: evaluation pipelines rigorous enough that "the policy works" is a proof, not a vibe. Run experiments across modalities — LLM agents, sim-to-real robot policies, game agents — on serious compute with no approval chain between idea and experiment. Publish and open-source when it helps the field. We build on open research and give back. ## What we\'re looking for - **PhD in RL, ML, robotics, or adjacent field strongly preferred** — or a research track record that makes the degree question boring. - First-author-level depth in at least one of: agentic RL for LLMs, RLHF/RLAIF, reward modeling, offline RL, or sim-to-real. - Strong engineering. Research here runs in production infrastructure, not notebooks. PyTorch or JAX fluency required. - Intellectual honesty bordering on allergy: every claim we make ships with proof, and you should like it that way. ## Nice to have - Experience with reward hacking/specification gaming research. - LLM post-training at scale (GRPO/PPO-family, large-batch async RL). - Built a benchmark or eval harness people actually use. ## What we offer - **$180K–$250K base** + meaningful founding equity. We benchmark against frontier labs because that\'s who we hire from. - Serious GPU budget and thousands of parallel rollouts from day one. - Founding-team leverage: your research decisions become the product. - Full health/dental/vision, relocation support, and the best hardware money buys. #J-18808-Ljbffr