PReMISE: A Framework for Auditing LLM Judge Rubrics

A novel framework named PReMISE (Policy Rubrics as Measurement Specifications for LLM Judges) has been launched to assess and audit the rubrics utilized by LLM judges. This initiative tackles the problem of ambiguous rubrics, which may favor well-articulated yet factually incorrect responses, such as those requiring answers to be 'helpful and factual.' PReMISE identifies policy-level rubric sets from pairwise human-preference data and evaluates any rubric set based on four criteria: structural adequacy, reliability, preference fit, and adversarial robustness. The findings indicate that no single raw rubric source achieves reliability, preference predictiveness, and adversarial robustness simultaneously, and high inter-rater agreement does not ensure low exploitability. PReMISE uniquely excels in applicability, specificity, and other metrics, aiming to enhance measurement specifications for LLM judges and promote more precise evaluations.