Two AI judges can disagree because they value style, detail or emphasis differently. Asking for a more elaborate rubric does not necessarily reveal those preferences.
PRISM's October 13 paper keeps evaluated content fixed while changing which attributes are emphasised. Human labels then train a correction from the judge's sensitivity profile. The authors report improved agreement with human rankings on summarisation and story evaluations. Read the research paper.
A judge's explanation is not a calibration test
Language models increasingly grade summaries, stories and other open-ended outputs. Their scores can disagree systematically with human judgements, and asking for a more elaborate rubric does not guarantee that the model reveals or corrects its implicit preferences. A judge may explain a score persuasively while remaining unusually sensitive to style, detail or the emphasis in its instructions.
PRISM uses the judge's reactions as observable data. It holds the evaluated content fixed and changes which attributes the judge is asked to emphasise or de-emphasise. The resulting score movements become features for a correction learned from human labels. The approach asks how the judge behaves rather than relying on its account of why it behaves that way.
Probe responses become a learned correction
The method's full name is Probe-based Residual Inference of Scoring Misalignment. A local sensitivity model motivates treating the response profile as a surrogate Jacobian. Ridge regression then learns how to correct the baseline score. The scoring protocol remains rubric-free in the sense that it uses overall scores without requiring separate criterion-level ratings.
That does not mean it is free of supervision. Human judgements supply the labels for calibration, and the choice of probes determines what kinds of sensitivity the correction can observe. A judge reacting strongly to a probe is not automatically biased: the experiment must preserve the content and identify whether that reaction predicts disagreement with the intended human evaluation.
Similar probes can support different judges
The main comparisons use Jev and Sol on SummEval summarisation and HANNA story evaluation. PRISM and a learned-baseline variant significantly improve agreement with human rankings over uncalibrated scores while remaining competitive with rubric-based methods in the authors' results. A controlled SummEval comparison finds similar agreement using rubric-derived quality probes and style probes, suggesting that reproducing every rubric criterion is not necessary for useful calibration.
The learned coefficients show both shared and opposing correction directions across judges. That is a reason to avoid treating one calibration vector as a universal repair. Different models can respond to similar interventions in different ways, and the correction is fitted to the particular judge and labelled setting.
Probe design needs controls of its own
The paper’s calibration setup compares probe responses against neutral clauses and uses held-out grouping when selecting regularisation. Those details address a basic danger: a fitted correction can learn accidental features of the calibration sample rather than a useful relation between the judge’s sensitivities and human disagreement.
The calibration needs fresh-content evaluation and retesting when the judge or human target changes. A new model version can alter probe responses even if its advertised identity stays similar, and a different rater population can change the intended ranking. Calibration is an empirical relationship among content, judge, probes and human labels. Its value comes from validating that relationship, not from treating the learned coefficient vector as a permanent description of fairness or quality.
The calibration has a domain and a price
Each probe adds inference calls. A comparison should account for those calls rather than presenting improved agreement as costless. Transfer across tasks, model versions and rater populations remains untested in the paper's stated limitations, and benchmark contamination cannot be ruled out. A correction fitted to stories should not be presumed suitable for scientific evidence or software correctness.
A reliable judge is not established by confidence, a detailed rubric or articulate justifications alone. Controlled interventions can reveal how its scoring shifts, and human labels can test whether those shifts help explain disagreement. PRISM turns those reactions into a calibratable signal while keeping the need for independent reference judgements explicit.
Questions, news tips or corrections? Contact [email protected].
