Senior Applied AI Researcher — Post-Training (RL) & Evaluation
2 days ago
Paris
About Lumos Lumos is a health and life-science AI evaluation company based in San Francisco. Our mission is to make health and life-science AI genuinely accurate and safe. We build the benchmarks, RL environments, and adversarial testing frameworks that frontier labs rely on to know whether their models can be trusted with science. We're scaling our Paris team to support new product launches and foundational lab partnerships — startup velocity, frontier-model stakes, in one of the most beautiful cities in the world. The Role You'll join our Evaluations function and build the systems that tell us whether medical models and agents are actually good — and actually safe — in a high-stakes healthcare setting. The role spans the full loop: designing RL environments, running post-training, and building the evaluations that close it back into the next model. These are genuinely open problems — how do you train and verify an agent to reason about medicine safely and reliably? — with real stakes and a lot of room to define the approach yourself. This is a hands-on, senior IC role: you build the environment and the harness, not just spec them. You'll work directly with our Head of Evaluations and partner closely with clinical subject-matter experts. • Design and ship RL environments for medical agents — tasks, tool interfaces, reward and verification logic., • Run post-training — RLHF / RLAIF / RLVR, reward modeling, preference-data pipelines, SFT, and RL fine-tuning., • Build evaluations that feed back into training — verifiers, graders, benchmarks, hard gates, and failure-mode analysis that directly shape the next model. Because you own both sides of the train↔eval loop, the signal you design shapes the next model directly rather than getting thrown over a wall. What You Will Do • Design and build RL environments for medical agents — defining tasks, tool interfaces, and the reward and verification logic that decides what "correct" means., • Own post-training of LLMs and agents end to end: RLHF / RLAIF / RLVR, reward modeling, preference-data pipelines, SFT, and RL fine-tuning., • Design case-specific verifiers, graders, and rubrics that capture medical correctness — in a domain that actually has ground truth, where reward design is a rich, hard, and largely unsolved problem., • Build evaluation frameworks — benchmarks, scoring dimensions, hard gates, red-teaming, and failure-mode analysis — and close them back into the training loop., • Train and evaluate multi-step agentic behavior: tool use, trajectories, and long-horizon decisions, not single-turn outputs., • Partner with clinical subject-matter experts to encode medical correctness into automated signal. What You Bring • 2+ years of direct, hands-on experience with model post-training — RLHF / RLAIF / RLVR, reward modeling, or RL fine-tuning loops. You've personally run these loops, not just read about them., • Experience with RL and/or environment design — agent environments, reward shaping, verifiable rewards. A major plus even if it isn't your whole background., • Experience designing evaluations for capable models — benchmark construction, LLM-as-judge / verifier systems, capability and safety evals., • Strong engineering ability — you can build the environment and the harness yourself. Python expected; we work across Python and TypeScript., • Scientific maturity — you care about reproducibility, statistical validity, and contamination, and you know the difference between a number that looks good and one that means some, • thing., • Strong written and spoken Enbglish. Bonus • Background in medicine, biology, or another rigorous scientific field; clinical exposure., • Publications or open-source contributions in RL, post-training, evaluation, or interpretability., • Experience with verifier-based or formal approaches to scoring correctness., • Prior work on safety-critical or domain-specific systems (healthcare, legal, security). Prior Experience This is a sourcing signal, not a filter — the underlying experience matters far more than the logo. Strong candidates often come from places where post-training, RL, and evaluation are first-class disciplines: • Frontier labs, • RL / environment & post-training — Prime Intellect, Mechanize, Scale AI, Surge AI, Snorkel, Contextual AI., • Eval / safety research orgs — METR, Apollo Research, Epoch AI, Redwood Research., • Applied AI teams that ran their own post-training and built internal RL environments and evals suites. Compensation • From €150,000 base salary, • Annual performance bonus, • 50k-110k options Work Authorization Candidates must be authorized to work in the European Union without sponsorship. Lumos does not sponsor work authorization for this role. Start date: Monday, July 27, 2026 (flexible) Equal Opportunities Lumos is an equal opportunity employer. We are committed to building a diverse and inclusive team and do not discriminate on the basis of race, color, religion, gender, gender identity or expression, sexual orientation, national origin, genetics, disability, age, or any other characteristic protected by applicable law. All qualified applicants will receive consideration for employment.
