Post-Training Applied Researcher

Post-Training Applied Researcher ABOUT BASETEN Baseten powers mission-critical inference for the world's most dynamic AI companies, like Cursor, Notion, OpenEvidence, Abridge, Clay, Gamma and Writer. By uniting applied AI research, flexible infrastructure, and seamless developer tooling, we enable companies operating at the frontier of AI to bring cutting-edge models into production. We're growing quickly and recently raised our $300M Series E https://www.baseten.co/blog/announcing-baseten-s-300m-series-e/, backed by investors including BOND, IVP, Spark Capital, Greylock, and Conviction. Join us and help build the platform engineers turn to to ship AI products. THE ROLE This role sits at the applied end of our post-training research efforts. You will work directly with stakeholders from the world’s fastest-growing AI companies to post-train open-source models that outperform frontier closed models on their specialised tasks. Your day-to-day is finding creative ways to extract signal from complex, domain-specific datasets and building the reward functions, environments, eval harnesses, and training pipelines that turn that signal into better models. The models you train ship to production and reach millions of users. We are looking for people with hands-on LLM fine-tuning and RL experience. Researchers who are excited by the prospect of shipping models into production, who can translate a customer's domain-specific requirements into an effective training curriculum, and who know when to be rigorous and when to iterate fast. RECENT RESEARCH - Dense, on-policy or both? https://www.baseten.co/research/dense-on-policy-or-both/ - Repeated kv cache for long-running agents https://www.baseten.co/research/repeated-kv-cache-for-long-running-agents/ - Distillation without the dark – replicating black-box on-policy distillation on Baseten https://www.baseten.co/research/distillation-without-the-dark/ RESPONSIBILITIES - Design and run post-training pipelines: SFT, GRPO, DPO, RLVR, reward function engineering, and synthetic data generation. - Build task-specific training environments and evals tailored to customer domains like healthcare, code generation, and legal, spanning multi-turn tool use, sandboxed execution, and agentic workflows. - Work directly with customers to translate production data into training signal, designing reward loops from real usage patterns and handling distribution shift. - Run and analyze training experiments end-to-end: diagnose reward hacking, importance sampling drift, and advantage estimation instabilities. - Publish findings at top venues and contribute to Baseten's open-source training libraries. QUALIFICATIONS - Hands-on experience training LLMs with reinforcement learning — demonstrated understanding of GRPO or PPO beyond recipe-level reproduction, including group advantage computation, clipped objectives, and KL penalty design - Strong intuition for reward engineering: the ability to distinguish between a reward that trains effectively and one that will exploit at scale - Experience building multi-turn agent environments with tool use, not limited to single-turn question-answering setups - Comfort working across the full pipeline from dataset construction through training, evaluation, and deployment - Experience with production ML systems. Preference for candidates who have closed a training–inference loop where production data feeds back into model improvement PREFERRED QUALIFICATIONS - Experience with RL training frameworks - Publications at NeurIPS, ICML, ICLR, focused on RL for LLMs, reward modeling, or alignment BENEFITS - Competitive compensation, including meaningful equity. - 100% coverage of medical, dental, and vision insurance for employee and dependents - Generous PTO policy including company wide Winter Break (our offices are closed from Christmas Eve to New Year's Day!) - Paid parental leave - Company-facilitated 401(k) - Exposure to a variety of ML startups, ... (truncated, view full listing at source)