LLM-as-a-Judge pipelines are increasingly used to evaluate AI-generated text, based on the assumption that judgments arise from reasoning over candidate responses with respect to a rubric. We show that this assumption warrants further scrutiny. Classifiers trained only on rubric text, without access to any evaluated response, achieve nontrivial predictive performance on judge outputs. This suggests that rubric formulations encode recoverable evaluative signals, allowing scores to be partially anticipated independently of model outputs. Finally, counterfactual perturbations reveal that judges often fail to reliably update their decisions when either the candidate response or the rubric criterion is reversed. Our findings raise concerns about the reliability of rubric-based LLM evaluation and highlight the need for further methodological study of automated evaluation via LLMs.
Figures & tables
Figure 1: Rubric Probe : Given a rubric alone, we train a classifier to predict the judge label, testing whether rubric text carries recoverable evaluative signal beyond chance.
Figure 2: Probe overview . The classifier receives only rubric text and predicts the LLM-as-a-judge output, while the evaluated answer, contextual metadata, and grounding information are excluded.
Figure 3: Per-metric comparison across probe datasets. PubMedBERT classifiers use rubric text to predict LLM-as-a-Judge labels. The left panel shows HealthBench results for Gemma, LLaMA, MedGemma, and MedLLaMA, while the right shows ResearchRubrics results for Gemma and LLaMA. The latter reflects the available candidate-model evaluations for this non-medical domain. Error bars denote standard deviation. The weighted classification (WC) and balanced subsampling (BS) cohorts are presented here to address class imbalance.
Figure 4: Output perturbation : The conversation and rubric are fixed while the candidate response is reversed. A counterfactually responsive judge should always flip its verdict. Instead, the expected verdict reversal occurs in only 37.7% of pairs, indicating that the judge often fails to track changes in answer-level rubric satisfaction.
Figure 5: Rubric perturbation : On HealthBench , the conversation and candidate response are fixed while the rubric criterion is reversed. A criterion-sensitive judge should therefore always flip its verdict. Instead, the expected reversal occurs in only 16.8% of pairs with Gemma and 32.2% of pairs with LLaMA, indicating poor sensitivity to changes in the evaluation criterion.
Figure 6: The UMAP projections of HealthBench-Eval-Probe and HealthBench-Hard-Probe
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
temperature
1.0
repetition penalty
1.15
top_p
0.95
num_return_sequences
1
Appendix
Table 1: Decoding hyperparameters used for all model queries. All the other parameters were set to default values.
Figure 8: Confusion matrices for rubric-only judge-label classification across the four evaluated models
Figure 9: Rubric-Only Classifiers Achieve Nontrivial Performance across Judge Models and Evaluation Metrics
Model
Metric
5-Fold CV (Mean ± SD)
Held-Out Test [95% CI]
Gemma
Accuracy
0.826 ± 0.006
0.826 [0.814, 0.837]
F1
0.682 ± 0.006
0.682 [0.671, 0.693]
Precision
0.687 ± 0.016
0.687 [0.656, 0.717]
Recall
0.704 ± 0.013
0.704 [0.678, 0.729]
Balanced Acc.
0.801 ± 0.006
0.801 [0.789, 0.812]
MedGemma
Accuracy
0.886 ± 0.004
0.886 [0.879, 0.893]
Appendix
Table 2: Five-fold cross-validation and held-out test-set performance of the rubric-only probe classifier. Confidence intervals are obtained using 10,000 bootstrap resamples of the held-out test set.
Figure 10: Cross-dataset transfer: within-split vs. cross-split balanced accuracy. Solid bars = within-split (train = test). Hatched bars = cross-split (train on one split, test on the other) Lower gap = better generalization.
Classifier
Gemma
MedGemma
LLaMA
Acc / Pre / Rec / F1
Acc / Pre / Rec / F1
Acc / Pre / Rec / F1
PubMedBERT
86.5 / 83.1 / 86.2 / 85.9
79.2 / 82.1 / 79.1 / 80.6
78.2 / 77.1 / 78.0 / 77.8
BERT
86.1 / 85.2 / 86.2 / 84.6
76.0 / 80.9 / 76.1 / 78.4
74.5 / 78.2 / 74.1 / 75.9
RoBERTa
85.2 / 81.5 / 90.2 / 85.3
78.4 / 83.7 / 78.2 / 81.1
75.9 / 78.7 / 75.4 / 76.6
DeBERTa-v3
84.6 / 84.5 / 86.7 / 84.9
70.5 / 83.4 / 71.6 / 74.3
74.2 / 78.3 / 74.7 / 75.1
Appendix
Table 3: Classifier comparison on HealthBench . Values are reported as accuracy, precision, recall, and F1 (%).
Response Model
HB -Eval
HB -Hard
n=4000
n=1000
Gemma
89.4
76.0
LLaMA
88.1
79.3
MedGemma
83.4
78.3
MedLLaMA
78.3
88.4
Appendix
Table 4: Majority-class accuracy (%) for each model on HealthBench -Eval HealthBench -Hard.
Outcome
Count
Percentage
Desired flip
186
37.7%
Wrong-direction flip
88
17.8%
Sticky satisfy
107
21.7%
Sticky violate
112
22.7%
Appendix
Table 5: Transition outcomes under counterfactual output perturbation. Percentages are computed over the corresponding 493 valid paired evaluations.
Outcome
Gemma
LLaMA
Desired flip
125 + 43 (16.8%)
252 + 70 (32.2%)
Sticky satisfy
8 (0.8%)
31 (3.1%)
Sticky violate
824 (82.4%)
646 (64.7%)
Appendix
Table 6: Transition outcomes under counterfactual rubric perturbation. The two desired flip directions are aggregated in the reported desired flip rate.
Figure 11: Lexical Cue Frequencies in Rubric Text. Mean frequency of selected binary lexical features across rubrics stratified by judge label.
Figure 12: Top Logistic Regression Weights by Judge Label. Rubric phrasing encodes a class-discriminative lexical signal
Figure 13: Rubric Word-Count Distribution by Judge Label. Label 0 rubrics are longer on average; distributions overlap substantially.
Figure 14: BERTopic Model results on HealthBench-Hard and HealthBench-Eval
Large language model (LLM)-as-a-judge research uses language models to score, compare, and verify AI outputs as scalable alternatives to human evaluation.