Judge disagreement clarifies rubrics but not human alignment
In plain words
When several language models use the same scoring guide, they may disagree because they read some rules differently. This paper introduces a way to find those disputed rules without using human ratings to choose them: an editor model clarifies clauses where judges disagreed.
On a test of news-summary ratings, the five judges became more consistent after the selected clauses were clarified. Their ratings did not detectably move closer to expert ratings, and the same consistency gain did not appear on the two other tested datasets.
The research
When judges disagree about an evaluation rubric, should an editor clarify the disputed clauses? Even if that makes judges agree more, will their ratings then better match human judgments?
The paper tests both outcomes separately, because consensus among judges does not by itself reveal which reading people intended.
The method uses judge disagreement to select rubric clauses for clarification, then separately checks judge consistency and agreement with human ratings.
The authors clarified clauses that judges had disagreed about, then compared ratings on separate evaluation items with ratings under the original rubric. On SummEval, five judges were more consistent with the edits.
A matched control found a similar gain only when the disputed clauses were targeted, not when the same number of unselected clauses were edited by the same procedure.
Yet the judges’ average ratings did not detectably track expert ratings more closely. The method therefore improved consensus in this test without showing that consensus better reflected human judgments.
The evidence covers three English datasets with existing human labels and a single edit to each selected clause. The authors transferred no criteria to code, so that stage was not tested.
Analysis
Deep
- Authors
- Sangdo Han, Moonsoo Kim
- License
- arXiv non-exclusive