Multi-Role Rubrics Improve LLM Evaluation and Reward Modeling

Evaluating and optimizing large language models (LLMs) for open-ended tasks requires reliable reward and preference signals. While rubric-based judges offer transparency by breaking down judgments into explicit criteria, existing automated rubric generators often rely on a single, generic evaluator. This can lead to "dimensional blind spots," where important aspects of human preference are overlooked. To overcome this, the researchers propose Multi-Role Rubric Generation (MRRG). This training-free and reference-free framework gathers evaluation criteria from various complementary roles, consolidating them into an auditable rubric-based scorer. This scorer can then be used for validating pairwise preferences and generating rewards for Reinforcement Learning with Verifiable Rewards (RLVR). Experiments demonstrate that MRRG consistently outperforms single-role rubric generation across different backbone models in preference validation benchmarks and provides a more robust reward signal for enhancing open-ended generation.