Rubricdropout
New preprint from my work at Scale AI: Rubric Dropout — a one-line fix that mitigates reward hacking in rubric-as-reward RL by randomly dropping rubric criteria at every training step. 🎲
New preprint from my work at Scale AI: Rubric Dropout — a one-line fix that mitigates reward hacking in rubric-as-reward RL by randomly dropping rubric criteria at every training step. 🎲