Description
At Harvey, we're transforming how legal and professional services operate by combining frontier agentic AI, an enterprise-grade platform, and deep domain expertise.
As a Research Engineer / Scientist, Reinforcement Learning, you'll own how we adapt and improve leading open-weight models for legal work, focusing on post-training efforts.
Role Overview You'll own the post-training recipe across SFT and RL, including reward design and verification for challenging, semi-verifiable legal tasks. You'll help set the research direction and carry ideas from experiments through training runs to models used by customers.
Responsibilities
- Own and pursue a research agenda to improve model capability on long-horizon legal reasoning and agentic workflows.
- Design and run post-training experiments across SFT and RL, building on leading open-weight models.
- Develop reward functions and verifiers for semi-verifiable tasks, including rubrics, learned preference models, and expert preference data, and study where they break.
- Run and debug training at scale, addressing divergence, reward hacking, and instability.
- Build and use rigorous evals to track improvements and decide what ships.
- Design, implement, test, and debug code across our research stack.
Requirements
- Hands-on experience post-training LLMs with SFT and RL, including preference optimization, and direct ownership of training runs and results.
- Deep ML fundamentals and experience getting distributed training runs working reliably.
- Strong engineering skills in Python, with production-quality code and end-to-end system ownership.
- Research experience in RL for language models or reward modeling, with the ability to move quickly from experiments to shipped improvements.
Compensation $230,000 - $340,000 USD per year