- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper asks whether LLM examiners’ explanations of their grades reflect what actually drives their grading. It adapts a method from human judgment research: infer a judge’s implicit priorities from its decisions, then compare them with the judge’s own account. Using synthetic answers to a real law exam, several LLM examiners grade the answers holistically and justify each grade. Their stated priorities are then compared with weights inferred statistically from their grading behavior, using rubric-based quality scores. The authors find that the examiners grade consistently and agree with one another, but their explanations match their behavior only partially. The models tend, for instance, to understate how much the sheer number of cited cases influences their grades.
Critical strengths
A principled, transferable framework for testing LLM explanations: Importing lens-model and policy-capturing methodology gives the field a structured, black-box way to compare what a judge says with what its decisions reveal.
Critical weaknesses
-
Unvalidated, non-independent measurements. The cue scores come entirely from LLM sub-judges, four of which are the examiner models themselves. No human expert validates these scores, the synthetic answers, or the justification coding. The self-report stage also shows examiners their own justifications before eliciting weights, so the agreement between the two is largely built into the design. These measurements underpin every comparison in the paper.
-
No baseline or statistical support for the self-insight claims. Each correspondence is a Spearman correlation over only ten dimensions, reported without confidence intervals, yet the paper ranks models by self-insight. More importantly, the self-reports may simply echo the rubric’s stated emphasis (which foregrounds ECHR jurisprudence) rather than reflect each model’s own grading. A simple test with existing data is missing: does each model’s self-report match its own behavioral profile better than other models’ profiles? Without such baselines, the headline claim that explanations are informative cannot be separated from a rubric-prior artifact.
Questions for authors
- Does each examiner’s self-report correspond more closely to its own behavioral profile than to the other examiners’ profiles? And how do self-reports from a model that never graded, given only the rubric, compare? If own-model correspondence is no better than these baselines, the results reflect the rubric’s emphasis rather than self-insight.
- How do the results change when each examiner’s own sub-judge scores are excluded from the cue measurements? Substantial shifts in R² or in the behavioral profiles would indicate that shared-model artifacts drive the findings.
- Were the sub-judge scores, the synthetic answers, or the justification coding validated by human experts, even on a subset? Did the sub-judge scores track the intended quality targets? Without such evidence, the behavioral policies may rest on mismeasured cues.
- What are the confidence intervals for the ten-point Spearman correlations? If the intervals overlap substantially, the model-to-model comparisons of self-insight are not supported.
- Were self-reports elicited without showing the examiner its own justifications? If not, the agreement between self-reports and justifications may be an artifact of the design rather than evidence of consistent introspection.
2 comments
Response to Reviewer GXZA: Questions for Authors
Question 1: Does each examiner’s self-report correspond more closely to its own behavioral profile than to the other examiners’ profiles? And how do self-reports from a model that never graded, given only the rubric, compare? If own-model correspondence is no better than these baselines, the results reflect the rubric’s emphasis rather than self-insight.
Response:
We conducted a cross-examiner comparison of each examiner's self-reported importance with the behaviorally recovered cue utilization policies of all four examiners (Appendix D, Table 16).
Only two examiners showed their highest correspondence with their own behavioral profiles. Mean within-examiner correspondence was 0.534, compared with 0.484 across examiners.
These findings highlight the value of our framework: it allows us to identify when an examiner's stated priorities correspond to its observed grading behavior and when they do not. Our objective is to assess this correspondence, rather than assume that examiners accurately describe their own judgments. We have clarified this distinction in Section 3.3 and included the full comparison in Appendix D, Table 16.
Question 2: How do the results change when each examiner’s own sub-judge scores are excluded from the cue measurements? Substantial shifts in R² or in the behavioral profiles would indicate that shared-model artifacts drive the findings.
Response:
Each examiner's own sub-judge scores were already excluded when calculating its averaged quality measurements. Consequently, the reported behavioral profiles are based on assessments from the four remaining models, avoiding direct overlap between the examiner and its sub-judges.
As a sensitivity check, we repeated the analysis using all five sub-judges. The recovered behavioral profiles remained highly similar ($\rho$ = 0.952–-0.988), with changes in $R^2$ no greater than 0.0072. These results indicate that our behavioral importance estimates are robust to the sub-judge aggregation choice. We have clarified the procedure in Section 2.3 and included the sensitivity results in the Appendix D Table 16.
Question 3: Were the sub-judge scores, the synthetic answers, or the justification coding validated by human experts, even on a subset? Did the sub-judge scores track the intended quality targets? Without such evidence, the behavioral policies may rest on mismeasured cues.
Response:
The ten quality dimensions were derived from the official grading guidelines and assessed using predefined scoring criteria. The synthetic answers were generated with intended quality targets, while justification coding followed a detailed codebook with explicit coding rules.
Our objective was to examine LLM judgment patterns rather than establish objectively correct grades. Although independent human-expert validation was not conducted, the standardized assessment procedure provides a consistent basis for comparing the examiners. For our purpose, the key consideration is whether the measurements capture the intended rubric dimensions, rather than whether they reproduce expert grading judgments.
Question 4: What are the confidence intervals for the ten-point Spearman correlations? If the intervals overlap substantially, the model-to-model comparisons of self-insight are not supported.
Response:
We calculated $95\%$ confidence intervals for all twelve Spearman correlations. The intervals are relatively wide, reflecting the limited number of dimensions ($n=10$). We therefore interpret the correlations descriptively and avoid drawing definitive conclusions about differences in self-insight between examiners. The confidence intervals have been added to the appendix D Table 17.
Question 5: Were self-reports elicited without showing the examiner its own justifications? If not, the agreement between self-reports and justifications may be an artifact of the design rather than evidence of consistent introspection.
Response:
The self-reports were intentionally elicited after the examiners received their previous grades and justifications, allowing them to reflect retrospectively on their completed evaluations. However, our primary focus is the correspondence between reported priorities and observed grading behavior, rather than agreement between the two verbal accounts. We therefore do not interpret self-report–justification correspondence as independent evidence of self-insight and have clarified this distinction in the revised manuscript in Appendix F.
Response to Reviewer GXZA: Critical Weakness
Critical Weakness 1: Unvalidated, non-independent measurements. The cue scores come entirely from LLM sub-judges, four of which are the examiner models themselves. No human expert validates these scores, the synthetic answers, or the justification coding. The self-report stage also shows examiners their own justifications before eliciting weights, so the agreement between the two is largely built into the design. These measurements underpin every comparison in the paper.
Response:
Critical Weakness 2: No baseline or statistical support for the self-insight claims. Each correspondence is a Spearman correlation over only ten dimensions, reported without confidence intervals, yet the paper ranks models by self-insight. More importantly, the self-reports may simply echo the rubric’s stated emphasis (which foregrounds ECHR jurisprudence) rather than reflect each model’s own grading. A simple test with existing data is missing: does each model’s self-report match its own behavioral profile better than other models’ profiles? Without such baselines, the headline claim that explanations are informative cannot be separated from a rubric-prior artifact.
Response: