Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning.

arXiv:2606.04923v3 Announce Type: replace-cross Abstract: Rubric-based reinforcement learning (RL) uses an LLM-as-a-Judge (LaaJ) to score model outputs according to rubrics as rewards. However, policy models may exploit latent biases in the judge, leading to r