PReMISE: Policy Rubrics as Measurement Specifications for LLM Judges
This work is significant for researchers and practitioners using LLM judges, as it provides a method to create more robust and reliable evaluation rubrics, addressing the problem of inconsistent and exploitable LLM judge performance.
This paper addresses the issue of LLM judge scores being highly dependent on the rubrics used. It introduces PReMISE, a framework that discovers policy-level rubric sets from human-preference data and audits existing rubric sets for structural adequacy, reliability, preference fit, and adversarial robustness. PReMISE improves judge accuracy from 65.0% to 68.6% and reduces exploit response high scores from 46.4% to 36.0%.
LLM judges are increasingly used to evaluate open-ended responses, but their scores depend strongly on the rubrics that condition them. A vague rubric asking for a response to be ``helpful and factual'' can reward polished answers that invent facts or violate user intent. We treat reusable rubrics as measurement specifications: changing the rubric changes the response quality measurement induced by a fixed judge. We introduce PReMISE, a framework that, given pairwise human-preference data, (i) discovers a policy-level rubric set, and (ii) audits any rubric set under LLM-judge use along four axes: structural adequacy, reliability, preference fit, and adversarial robustness. Across rubric sources no raw source is simultaneously reliable, preference-predictive, and adversarially robust; and high inter-rater agreement does not imply low exploitability. PReMISE is the only rubric source to score non-trivially on applicability, specificity, and effective dimensionality simultaneously. We contribute two audit-targeted repair operations: preference-rank selection raises judge accuracy on paired responses from $65.0\%$ to $68.6\%$, competitive with the strongest rubric-discovery baselines and leading on two of three judges in our cross-judge sweep; reliability-constrained refinement reduces the rate at which exploit responses receive high scores from $46.4\%$ to $36.0\%$ with little change in inter-judge agreement ($α{=}.531\to.519$).