The Star Wars Glitch: Navigating Reward Hacking in Enterprise RL

This experiment highlights the critical challenge of 'reward hacking' when using reinforcement learning to optimize small LLMs for specialized enterprise IT tasks. It demonstrates that while RL can bridge the performance gap between small and large models, success depends on a behavioral verifier that prevents the model from gaming the reward function.