Rubric content drives large language model grading outcomes more than judge choice
Rubrics Carry the Standard: Shapley Attribution for Checklist-Based LLM Judges
Computation and Language
Summary
It turns out that when AI models grade open-ended tasks, the specific checklist or rubric they use influences the pass rate far more than which AI model is doing the judging. The authors found that trying to have the AI improve its own rubric often leads to changes that don’t reflect the true task goals, making the rubric worse. They developed a mathematical way to measure each rubric check's impact, which helps identify checks that uniquely catch errors or cause unfair failures. Using this, they can prevent rubric changes from drifting and keep grading consistent across different AI judges.
What this means in practice
- •For machine learning engineers: Control and monitor rubric changes to maintain consistent grading standards across AI model evaluations of open-ended tasks.
- •For ai quality assurance teams: Analyze individual rubric checks for their unique contributions to error detection and reduce false failure rates in automated grading systems.
Authors
Saeid Asgari, Rui Ying, Sara Malvar, Rakshanda Agarwal, Bruce Sun, Claire Wang, Leonardo de Oliveira Nunes, Ranveer Chandra
Abstract
Checklist rubrics applied by large language model (LLM) judges have become a standard instrument for evaluating open-ended outputs and for rewarding them during training. Two assumptions surround this practice: that leniency or strictness is a property of the judge model, and that a rubric can be improved by asking a model to review it. We test both in a controlled grading task with planted flaws and on two human-labeled datasets, SummEdits and RAGTruth. Holding the judge and the outputs fixed, rubric content moves the pass rate of correct outputs from 2% to 100%, whereas changing the judge model or the rendering of the same checks moves it by a few points. Model self-review then drifts the rubric toward the reviser's own standard: it adds requirements the task never stated and deletes checks that human labels had taught, with three reviser models and on all three datasets. To measure and control this, we show that under all-must-pass aggregation the Shapley value of each check has an exact closed form, which yields per-check catch credit, false-fail blame and unique catches in linear time. Attribution shows that self-review deletes the checks that do unique work and keeps the checks that cause false fails. A change-control gate that combines these quantities with a provenance freeze on label-derived checks holds nearly every drifting revision chain at its starting quality, with identical decisions under three judge models.