RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains

Researchers proposed RUBRIC-ARROW, an alternating reward-modeling framework for LLM post-training.

It trains a rubric generator and a rubric-conditioned judge, then uses pairwise preference data in reinforcement learning.

The method replaces tie-prone pointwise scoring with probability-based rubric scoring to better separate answers.

Experiments showed stronger reward-modeling performance and improved downstream post-training results, including GRPO.