RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains
For LLM post-training in subjective domains, this method improves reward modeling by reducing ties and leveraging rubrics without relying on frontier LLMs.
RUBRIC-ARROW addresses the challenge of pointwise reward modeling in subjective, non-verifiable domains by jointly training a rubric generator and a rubric-conditioned judge using only pairwise preference data, achieving competitive accuracy and consistent gains in downstream policy post-training.
Pointwise reward modeling offers critical signals for LLM post-training, yet struggles with absolute scoring in subjective, non-verifiable settings. Rubric-based methods address this by decomposing evaluation into explicit criteria, but existing approaches typically depend on frontier LLMs and suffer from ties caused by hard Boolean aggregation. We present RUBRIC-ARROW, an alternating framework that jointly trains a rubric generator and a rubric-conditioned judge, with its RL stage using only pairwise preference data. Our method couples a probability-based scoring rule that reduces ties with phase-specific preference-based rewards and an alternating GRPO scheme that together train the pointwise evaluator. Extensive experiments show that RUBRIC-ARROW achieves competitive reward-modeling accuracy and yields consistent gains for downstream policy post-training.