Member of Technical Staff - RL Algorithms

RL has become the de-facto method of post-training LLMs. We are limited by the sample efficiency of the current policy gradient algorithms in use today, and are looking for a talented researcher to weave together pre-LLM and post-LLM approaches to learning from experience.

• Develop new RL algorithms for post-training language models.
• Adapt ideas from pre-LLM reinforcement learning, such as model-based RL, temporal abstraction, and value-based learning, to modern LLM and agentic settings.
• Establish empirical baselines and evaluation protocols for measuring sample efficiency, robustness, generalization, and reward exploitation in LLM RL.
• Analyze failure modes of RL-trained models, including reward hacking, mode collapse, over-optimization, exploration failures, and distribution shift.
• Collaborate with researchers working on environments, evals, interpretability, reward modeling, and infrastructure to turn algorithmic ideas into reliable training systems.