The RL Fine-Tuning Playbook: CoreWeave's Kyle Corbitt on GRPO, Rubrics, Environments, Reward Hacking

Reinforcement learning (RL) offers a superior path for optimizing LLM performance compared to supervised fine-tuning (SFT) by working within a model's existing weight distributions, thereby reducing catastrophic forgetting. While SFT often overrides established pathways, RL optimizes for specific outcomes by reinforcing critical decision points in reasoning traces. The GRPO algorithm, which scales RL by running multiple parallel rollouts and calculating group-relative advantages, has become an industry standard for improving reasoning capabilities. Despite concerns regarding reward hacking, iterative rubric development and auxiliary LLM judges effectively mitigate these risks. Compute remains the primary constraint for scaling, though recursive self-improvement loops are already accelerating development. For enterprises, RL-driven fine-tuning provides significant latency and cost advantages, particularly for specialized agentic tasks where off-the-shelf frontier models are either too slow or too expensive to deploy at scale.