PaperlayerView exact revision on arXiv

Source-linked AI summary

JudgeProfile: Understanding and Steering Subjectivity in LLM Judges

Qi Cao, Kangning Liu, Xuan Kan, Shunwen Tan, Yang Pei, Dake Chen, Yatai Ji, Zixuan Ye, Yuanpeng Tu, Daniel Li, Junbiao Tang, Pengtao Xie, Zihao He

arXiv:2609.36705v1cs.AI

TL;DR

LLM judges can disagree about which response is better even when neither is objectively wrong, and an overall preference alone does not show why. JudgeProfile separates attribute perception from prioritization and studies both using SubjectiveSet. Reweighting improves held-out agreement with reference labels, while the findings show that judges can share perceptions but differ in how they value them.

  • Problem

    For open-ended responses with no single objectively correct judgment, it remains unclear what drives LLM judges’ differing preferences.

    Source: §1 · p. 1

  • Method

    JudgeProfile separates attribute-level perception from prioritization, using SubjectiveSet’s 50,013 response pairs, 21 LLM judges, and 87 attributes to study both.

    Source: §1 · p. 1

  • Results

    Reweighting raises average held-out agreement with reference labels from 66.48% to 71.97%, while judges also agree on attributes in 85.9% of analyzed comparisons despite differing overall choices.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    Examining attribute judgments and fitted weights can help explain why judges disagree and adapt their overall decisions without updating the language model.

    Source: §6 · p. 10

    A single weight vector per source cannot capture query-dependent trade-offs or interactions among attributes, which may help explain why full-data fine-tuning surpasses reweighting on three of four judges.

    Source: §6.1 · p. 10

Abstract

from arXiv · show

LLM judges are inherently subjective, often favoring different responses in pairwise comparison when neither option is objectively wrong. To study this subjectivity, we introduce JudgeProfile, a framework that dissects LLM evaluation into perception (how a judge compares two responses across specific attributes like clarity, correctness, and detail) and prioritization (how much each attribute influences the final choice). We curate SubjectiveSet, a dataset of 50,013 response pairs from 17 public data sources, evaluated by 21 LLM judges across 87 attributes. We find a hidden consensus in perception: judges frequently agree on attribute judgments even when their overall choices diverge. Building on this separation, we first characterize each judge's prioritization using attribute weights estimated from its own overall choices. These weights differ across judges even when estimated from the same attribute judgments. We then learn new weights from reference labels to adapt their decisions to a target evaluation standard. Reweighting perceived attributes improves average held-out agreement with reference labels from 66.48% to 71.97%, outperforming fine-tuning and rubric prompting. Our findings show that understanding and steering the subjectivity of LLM judges requires attention not only to what they perceive, but also to how they prioritize it.

1 Introduction

JudgeProfile separates LLM judgments into attribute-level perception and the prioritization that turns those assessments into overall choices. The paper finds that judges often share attribute judgments but differ in their priorities, and that reweighting can improve agreement with target labels without changing the model.

1 Introduction

  • JudgeProfile separates perception—comparing responses on attributes such as clarity, correctness, and detail—from prioritization, which weights those attributes in an overall judgment.Subjectivity may arise from differing perceptions, differing priorities, or both.Source: §1 · p. 1
  • SubjectiveSet contains 50,013 response pairs from 17 public data sources, evaluated by 21 LLM judges across 87 attributes.Each attribute is evaluated separately in both response orders, yielding 182.7 million inference calls.Source: §1 · p. 1
  • Even when overall choices differ, judges agree on which response has more of a given attribute in 85.9% of comparisons analyzed.The paper calls this pattern hidden consensus; a human study shows the same pattern.Source: §1 · pp. 1–2
  • Judges’ fitted attribute weights differ even when estimated from the same attribute judgments, identifying prioritization as another source of subjectivity.The weights remain reproducible within each judge across data splits.Source: §1 · p. 2
  • Reweighting raises average held-out agreement with reference labels from 66.48% to 71.97%, improving results for all 21 judges.With 256 labels per source, it outperforms LoRA fine-tuning on all four tested judges; with six attributes, it beats rubric prompting on three judges and performs similarly on the fourth.Source: §1 · p. 2
  • The framework and dataset support studying subjectivity in both LLM judgments and a smaller set of human judgments.The paper also proposes lightweight steering by learning attribute weights without updating the language model.Source: §1 · p. 2

2 Subjectivity in LLM Judges

The section defines subjectivity as variation without a unique objective ground truth and distinguishes attribute-level perception from the prioritization behind overall preferences. It then describes a large-scale study designed to compare those judgments across LLMs and humans.

2.1 What is subjectivity?

  • Subjectivity refers to variation in judgments for which there is no unique objective ground truth; disagreement alone does not establish it when external evidence can settle a factual claim.Evaluative qualities such as whether a response is interesting or sufficiently detailed may lack a single objectively correct answer.Source: §2.1 · p. 2

2.2 Perception and prioritization.

  • Perception captures how a judge compares responses on individual attributes, while prioritization captures how much those differences matter to its overall preference.Separating attribute comparisons from holistic verdicts tests whether disagreement reflects different perceptions, different priorities, or both.Source: §2.2 · p. 3

2.3 Studying subjectivity at scale.

  • For each response pair, 21 LLM judges provide overall preferences and separate judgments across 87 attributes; evaluating every attribute independently in both response orders yields 182.7 million attribute-level inference calls.SubjectiveSet includes preference-labeled data based on human or AI preferences and benchmark-labeled data based on benchmark evaluation criteria; LLM disagreement selects pairs for closer attribute-level analysis.Source: §2.3 · p. 3
  • In the home-workout example, judges agree response B is more specific and response A more insightful, yet disagree about which response is better overall.Source: Figure 4 · p. 4

3 Perception and Subjectivity

JudgeProfile separates subjectivity into perception of response attributes and prioritization of those attributes in overall judgments. Attribute judgments often agree even when judges choose different winners.

3 Perception and Subjectivity

  • The authors call agreement on attributes despite differing overall judgments “hidden consensus.”They report observing this pattern among LLM judges and in a small human study.Source: §3 · p. 4

3.1 Evidence 1: Attribute judgments agree even when judges choose different winners.

  • The section compares attribute agreement when judges select the same versus different overall winners.Figure 3’s comparison uses attribute judgments alongside overall judgments.Source: §3.1 · pp. 4–5
  • 91.9% mean attribute agreement accompanies the same winner, versus 85.9% when 19 judges choose different winners.The analysis excludes two judges with severe position bias in their attribute judgments.Source: §3.1 · p. 4

21 LLM judges

Across LLM and human disagreement cases, attribute judgments remain more aligned than overall choices. However, agreement varies substantially across attributes, including some judged important.

21 LLM judges

  • Among 19 judges split 8–11 on a home-workout winner, all found response B more specific, while 18 consistently found A more insightful.Their attribute judgments agreed even though their overall choices differed.Source: §21 · p. 5

3.2 Evidence 2: Model and human disagreement cases show the same pattern.

  • 84.3% attribute agreement contrasts with 49.8% overall agreement on pairs selected for disagreement among 21 LLM judges.Table 1 reports pairwise agreement on disagreement cases.Source: §3.2 · p. 5, Table 1 · p. 5
  • On 21 human disagreement cases, seven raters show 92.8% attribute agreement and 57.6% overall agreement.The authors report this as a similar hidden-consensus pattern in human judgments.Source: §3.2 · p. 5

3.3 Discussion: Attribute judgments can still be subjective.

  • When judges choose different winners, agreement is 66.0% for final-answer correctness and 64.2% for factual accuracy, versus 91.5% for harmfulness.The listed attributes come from the highest importance quartile; high average agreement hides substantial disagreement on some important attributes.Source: §3.3 · p. 5

4 Prioritization and Subjectivity

JudgeProfile examines whether judges’ overall disagreements reflect different attribute judgments or different prioritization. It finds distinct, reproducible priorities and shows that learning new weights can improve alignment with reference labels.

4.1 Evidence 1: Shared attribute judgments do not guarantee the same winner.

  • Winners still differ in 14.5% of eligible cases where judges agree on all observed stable shared attribute directions.The comparison requires definite overall judgments and at least 20 shared attributes with definite, order-consistent judgments.Source: §4.1 · p. 6

4.2 Evidence 2: Fitted priorities differ across judges and are reproducible within judges.

  • Fitted priorities are reproducible within judges (0.911–0.986 reliability) but differ across judges (mean correlation 0.609).Even with the same attribute representation, weights differ (correlation 0.575), with similar prediction accuracy (0.820 versus 0.822).Source: §4.2 · p. 6

4.3 Experiment 1: Reweighting fixed attribute judgments.

  • 66.48% to 71.97%: reweighting raises mean held-out accuracy across 21 judges, with gains across datasets.The reported mean gain remains 3.99 percentage points across the other 16 datasets when RM-Bench is excluded.Source: §4.3 · p. 6, Figure 8 · p. 7

4.4 Experiment 2: Reweighting versus fine-tuning.

  • At 256 labels per source, reweighting outperforms LoRA SFT on all four tested judges; for three, 32 labels already exceed SFT-256.On three tested distributions, gains flatten after 128–256 labels; the authors suggest fixed attribute judgments narrow the learning problem, while more labels may help SFT change more than priorities.Source: §4.4 · p. 7, Table 2 · p. 8

4.5 Experiment 3: Reweighting versus rubric prompting.

  • With six attributes in each method, reweighting is more accurate on three judges and trails by 0.07 percentage points on Qwen3.5-27B.Rubrics name criteria but leave the judge to determine their relative importance; reweighting directly fits attribute contributions.Source: §4.5 · pp. 7–8, Table 3 · p. 9

4.6 Discussion: Which attributes matter varies by the judge and dataset.

  • Gemma-4-31B favors instruction adherence over correctness, while Gemma-3-27B favors empathy over correctness; verbosity’s association with preference reverses across two datasets.Verbosity is positively associated with preferred responses in Nectar and negatively associated in PPE Best-of-K.Source: §4.6 · p. 8, Figure 10 · p. 8

5 Related Work

Prior work establishes LLMs as evaluators and studies their biases, attributes, preferences, and methods for adaptation. JudgeProfile builds on this work by examining perception and prioritization together.

5.1 LLM-as-a-Judge and subjectivity.

  • MT-Bench and G-Eval establish LLMs as evaluators, while later work studies length and position biases and combines diverse judges to reduce reliance on one model.The cited examples include Zheng et al. (2023), Liu et al. (2023), Dubois et al. (2024), Shi et al. (2025), and Verga et al. (2024).Source: §5.1 · p. 9

5.2 Attribute judgments and interpretable weights.

  • Prior work uses attributes and weights to explain preferences, including FLASK, HelpSteer, VibeCheck, WIMHF, and ArmoRM.JudgeProfile uses a shared attribute inventory to study agreement across LLM judges and how shared attribute judgments can coexist with differing overall judgments and fitted weights.Source: §5.2 · p. 9

5.3 Adapting LLM judges.

  • Prior adaptation approaches include fine-tuning evaluators on generated judgments or custom criteria and controlling evaluation with rubrics [Zhu et al., 2025] [Kim et al., 2024b,a] [Roy et al., 2026] [Chen et al., 2026].JudgeProfile tests changing attribute weights while keeping recorded attribute judgments fixed, comparing against fine-tuning and rubric prompting.Source: §5.3 · p. 10

6 Conclusion

JudgeProfile distinguishes agreement on response qualities from disagreement about how those qualities shape overall preference. The conclusion argues that evaluation should make attribute priorities and trade-offs explicit, while noting limits of a single linear weight vector.

6 Conclusion

  • JudgeProfile separates perception of response attributes from prioritization, and finds that judges can agree on qualities yet differ in overall choices and fitted weights.The conclusion also reports substantial attribute consensus in human judgments and that reweighting can improve agreement with target labels without updating the language model.Source: §6 · p. 10

6.1 Limitations and future work.

  • A single linear weight vector cannot capture query-dependent trade-offs or interactions among attributes, and full-data fine-tuning surpasses reweighting on three of four judges.The authors identify context-dependent or nonlinear weighting as a direction for future work.Source: §6.1 · p. 10

6.2 Broader implications.

  • Judge training and evaluation should specify not only preferred responses, but also which attributes to prioritize and how to resolve trade-offs.The paper states that explicit priorities make judge behavior easier to audit and adapt.Source: §6.2 · p. 10

B. Perception: Agreement and Robustness

The supplied passages describe curated subsets and cases used to examine response disagreement, attribute trade-offs, and the scope of LLM and human analyses. They also distinguish the sources and populations underlying those analyses.

A.2 SubjectiveSet: subset definitions and intended use.

  • 11,524 pairs across 17 sources meet the SubjectiveSet–LLM selection rule, while 38,489 pairs remain outside it.Source-specific selection rates range from 7.7% on RewardBench to 43.1% on IF-RewardBench; main results use the 21-judge definition.Source: §A.2 · p. 15
  • In the arithmetic example, judges favor A on humor (19/21) and B on rigor (18/21), so the preferred response is not higher on every attribute.The source reference label is B.Source: §A.2 · p. 16
  • In the pancake example, judges favor B on explanatory depth (21/21) but A on conciseness (18/21).The benchmark label does not separately rank the attributes.Source: §A.2 · p. 16

G. SubjectiveSet–Human: Complete Case Studies

  • The SubjectiveSet–Human collection contains 23 selected cases, including 21 model-disagreement cases and two controls, with seven raters evaluating 55 item–attribute cells.Nineteen cases meet the SubjectiveSet–LLM threshold; the set was curated to make response differences inspectable.Source: §A.2 · p. 16

A.3 Which data each analysis uses.

  • The LLM and human analyses illustrate a similar qualitative contrast but use different populations, attribute selection, and aggregation.The human aggregate pools eligible rater-pair comparisons, whereas LLM agreement averages eligible pairwise rates equally over judge pairs.Source: §A.3 · p. 17

A.4 Judges and analysis population

Across judges and sources, stable attribute judgments often agree even when judges choose opposite overall winners. The reported contrast depends on reliable, jointly resolved directions, and the selected human cases do not estimate prevalence in a random population.

A.4 Judges and analysis population

  • The 21-judge collection includes two models excluded from the main 19-judge agreement analyses because of severe position bias in attribute judgments.Both models choose position B about 80% of the time on order-balanced data, including after response swaps; they remain in collection-scale counts, 21-judge agreement analyses, and broad adaptation evaluation.Source: §A.4 · p. 17, Table 7 · p. 18

A.5 Attribute inventory and directional comparisons

  • The attribute task asks which response exhibits more of a specified characteristic; a direction toward A does not itself mean A is better.The evaluation uses 87 attributes grouped into 14 families.Source: §A.5 · p. 18, Table 8 · p. 19

A.8 Presentation order, inference scale, and coverage

  • 182.7 million attribute-level inference calls cover 87 attributes for each of 21 judges; holistic preference calls are additional and excluded from that count.Each attribute comparison is elicited in both response orders and mapped back to original response identities; the call count is not the number of stable or usable directions.Source: §A.8 · pp. 18–19

B.1 Order consistency and conditional agreement

  • 0.718–0.983 is the range of stable opposite-winner attribute agreement across the 19-judge pairs, with all 171 pairs above 0.5.Across distributions, opposite-winner agreement ranges from 0.817 to 0.897, with all 171 judge pairs above 0.5 in each distribution.Source: §B.1 · p. 22, §B.2 · p. 22

B.3 Matched items and single-read controls

  • 0.810 core-20 attribute agreement exceeds 0.767 overall preference agreement and 0.765 all-87 attribute agreement under the single-read protocol.Under stable filtering, the attribute-agreement gap is positive for all 171 judge pairs and all 19 per-judge averages; the all-attribute single-read advantage is absent.Source: §B.3 · p. 24, §B.4 · p. 24, Table 13 · p. 24

B.5 Reference-label difficulty control.

  • Agreement remains high when both judges depart from the reference label, using reference-label concordance as a proxy for difficulty.The analysis excludes 3,092 response pairs with tie reference labels and separates items where both judges match the reference, both depart, or disagree.Source: §B.5 · p. 24, Table 14 · p. 25

B.6 Human agreement and study scope

  • The 23 selected human cases show the qualitative contrast between preference and attribute judgments, but do not estimate its prevalence in a random population.Seven raters assessed 21 model-disagreement cases and two controls; possible anchoring from question order and the small control set further limit inference.Source: §B.6 · p. 25

C.1 Weight estimation and reproducibility

The paper estimates judge-specific attribute weights from each judge’s own stable attribute directions and holistic verdicts, then tests whether weight differences persist under shared representations. The reported fits reveal substantial cross-judge variation, while component swaps have important interpretive limits.

C.1 Weight estimation and reproducibility

  • Priority-model raw half correlations range from 0.837 to 0.973, while Spearman–Brown reliability ranges from 0.911 to 0.986.The models are L2-regularized logistic regressions fitted from each judge’s 87 stable directions to its own holistic verdicts.Source: §C.1 · pp. 25–26, Table 19 · p. 28

C.2 Refitting with shared attribute judgments

  • Refitting on shared majority attribute directions yields mean cross-judge weight correlations of 0.575 (Pearson) and 0.495 (Spearman), with prediction accuracy of 0.820 versus 0.822 using own perception.The shared representation uses the cell-wise majority direction across judges, with zero for a tied or absent majority.Source: §C.2 · p. 26, Table 17 · p. 27

C.3 Component swaps and their interpretation

  • The component-swap control reports in-sample accuracies of 0.822 with own perception and weights, 0.739 after a perception swap, and 0.799 after a weight swap.The authors caution that the swaps alter different objects and are not matched in perturbation size, so the drops measure sensitivity to replacements rather than relative causal contributions.Source: §C.3 · pp. 27–28

C.4 Estimation noise and source-specific fits

  • Observed cross-judge weight dispersion exceeds the null 95th percentile on 85 of 87 attributes, with a median observed/null-mean SD ratio of 2.16.Depth and friendliness are the two exceptions; weights are normalized by each judge’s mean absolute magnitude across attributes.Source: §C.4 · p. 28, Table 20 · p. 29

C.6 Judges assign different priorities to shared attributes.

  • All 19 judges assign positive weights to correctness and strategy quality, but their relative priorities differ.Gemma-4-31B weights instruction adherence above correctness, whereas Gemma-3-27B weights empathy above correctness.Source: §C.6 · p. 28

C.7 Reference-label associations depend on the dataset.

Reference-label associations for style can reverse across datasets, while fitted priorities vary by judge and two low-coverage judges warrant caution.

C.7 Reference-label associations depend on the dataset.

  • 79% for verbosity on Nectar versus 27% on PPE Best-of-K, while conciseness reverses at 34% versus 78%.Correctness predicts the reference winner above 50% in all 17 datasets; alignment therefore depends on the target and distribution.Source: §C.7 · p. 29

C.8 Attribute-priority display details.

  • Shared-representation priority correlations are 0.765 for perception and 0.645 for priorities across 171 judge pairs.Figure 14 displays normalized logistic coefficients for 19 judges on the same 13 selected attributes.Source: Figure 13 · p. 30, Figure 14 · p. 30

C.9 Sensitivity to the judge set.

  • The 21-judge sensitivity check preserves 12 shared axes, and the Gemma instruction-adherence and empathy contrasts hold in both displays.Qwen3.5-0.8B and Qwen3.5-2B have low stable coverage (16.5% and 10.6%), so their fitted priorities warrant greater caution.Source: §C.9 · p. 31, Figure 16 · p. 32

C.10 Source-specific associations are not fitted priorities.

  • Source-specific verbosity and conciseness associations reverse, and the pooled column can hide those reversals.RM-Bench is separated because its sampled chosen-style-0/rejected-style-2 construction strongly opposes length and formatting to its target labels.Source: §C.10 · p. 31, Figure 15 · p. 31

C.11 Selected coefficients and per-judge rankings

  • Fitted importance correlates negatively with opposite-winner agreement across 87 attributes (Pearson −0.253; Spearman −0.379), although harmfulness and hallucination retain high consensus.Coefficients are normalized within each judge; fitted importance uses each judge’s own attribute representation and verdicts.Source: §C.11 · p. 35

D.1 Fitted attribute importance.

The section defines fitted attribute importance from logistic coefficients and relates it to judges’ agreement on attributes when their overall winners differ.

D.1 Fitted attribute importance.

  • Fitted importance is the absolute logistic-regression coefficient predicting a judge’s overall verdict from 87 attribute judgments.The model is fit to each judge’s own overall verdicts.Source: §D.1 · p. 37
  • Coefficient scales are normalized within each judge before averaging coefficient magnitudes.The appendix uses 21 judges and 210 judge pairs under the order-consistent-direction protocol.Source: §D.1 · p. 37
  • Across 87 attributes, fitted importance correlates negatively with opposite-winner agreement (Pearson −0.253; Spearman −0.379), while harmfulness and hallucination retain high consensus.Agreement is conditioned on joint stable resolution, and coverage reports how often both judges resolve an attribute in the corresponding winner stratum.Source: §D.1 · p. 37

E Overall Disagreement Without Attribute Conflict

The paper measures overall disagreement despite agreement on resolved attributes, then evaluates reweighting across judges and datasets, noting that label-source groups are not pure objective-versus-subjective categories.

E Overall Disagreement Without Attribute Conflict

  • Even without conflict among measured directions, overall disagreement varies by source from 8.0% on RewardBench to 27.8% on IF-RewardBench.Unresolved attributes, difference magnitudes, unmeasured qualities, and noise may also contribute to this residual.Source: E Overall Disagreement Without… · p. 41, Table 35 · p. 42

F Reweighting and Adaptation

  • Reweighting fits dataset-provided training labels to each judge’s fixed attribute comparisons; target labels may reflect human preferences, model judgments, or verifiable answers.The evaluation covers 21 judges and 17 datasets, yielding 357 judge–dataset combinations.Source: §F.1 · p. 43

F.1 Evaluation, aggregation, and label sources

  • Judge-level means weight datasets equally, dataset-level means weight judges equally, and the overall macro gain does not weight datasets by test-set size.The preference-only aggregate covers 11 sources; excluding RM-Bench leaves 16 sources.Source: §F.1 · p. 43

F.2 Precision.

  • Detailed matrix accuracies and gains are rounded to one decimal, so aggregates may differ by 0.1 points.Source: §F.2 · p. 43

F.3 Grouping by label source.

  • The source-level partition is operational, not a pure subjective-versus-objective split; objective-only analysis of mixed benchmarks requires item- or subtask-level annotations.Some benchmarks include correctness-based comparisons, instruction-following targets, or mixed domains.Source: §F.3 · p. 43

F.5 Breadth of the improvement.

  • 66.48% to 71.97% average accuracy across 17 datasets, with all 21 judges improving; excluding RM-Bench, the remaining 16 datasets gain 3.99 points.Across the 11 preference-labeled datasets, accuracy rises from approximately 63.50% to 68.48%, a 4.98-point mean gain.Source: §F.5 · pp. 43–44, Table 36 · p. 44

F.6 Dataset heterogeneity and limitations.

This section examines how reweighting gains vary across datasets and uses diagnostic comparisons and case studies to explore the limits of rubric analysis. The cases show that agreement on response attributes can coexist with different overall preferences, shaped by factors such as humor, creativity, conversational intent, emphasis, prompt interpretation, and source limitations; response judgments alone do not establish factual or medical correctness.

F.7 RM-Bench comparison setting.

  • In RM-Bench’s 0:2 pairing, mean accuracy across 21 judges rose from 65.05% to 94.44% (+29.37 points), while longer-response win rate was 0.038.The pairing compares a plain chosen response with a more elaborate, formatted rejected response; excluding RM-Bench, the mean gain was 3.99 points across the other 16 datasets.Source: §F.7 · p. 44, Table 37 · p. 45

F.6 Dataset heterogeneity and limitations.

  • Mean gains were positive on 15 of 17 datasets, but ranged from +29.37 points on RM-Bench to −2.35 on JudgeBench, where individual gains spanned −16.2 to +16.2 points.HelpSteer2 was nearly unchanged (+0.11); RewardBench had a −0.18 mean gain. Outside RM-Bench’s 0:2 setting, gains were largest on preference-labeled sources, though PPE Best-of-K also benefited (+7.84).Source: §F.6 · p. 44, Table 37 · p. 45

F.11 Rubric construction and diagnostics.

  • With six attributes in both methods, reweighting exceeded rubric prompting by 1.75, 2.80, and 2.62 points on Qwen3.5-4B, Qwen3.5-9B, and Llama-4-Scout, respectively, but trailed by 0.07 on Qwen3.5-27B.Both methods used the same training-label budget and held-out test set; attribute selection and fitting used training data only.Source: §F.11 · p. 48

F.13 Separating changes in perception from changes in weights.

  • Weight alignment and linear explainability do not measure attribute-direction flips, so the comparison establishes performance and intervention control without identifying the cause of rubric effects.A direct diagnostic would re-elicit attributes under both prompts and measure flips among cells stable in both conditions, distinguishing off-target from explicitly targeted changes.Source: §F.13 · p. 48, §F.11 · p. 48

F.14 Context-dependent weights and controlled training.

  • A single source-level weight vector cannot express every query-dependent trade-off, which may help explain why full-data fine-tuning surpasses reweighting on three judges, though the cause is not identified.The proposed controlled follow-up varies target trade-offs on the same response pairs and tests whether explicitly stating the desired trade-off improves adaptation sample efficiency.Source: §F.14 · pp. 48–49

G.3 Case index.

  • Human and LLM attribute majorities matched on 54/55 comparisons, but overall majorities matched on only 13/22 cases with a human majority.The case set contains 21 model-disagreement cases and two controls; examples distinguish recognizing a quality from deciding its importance.Source: §G.3 · p. 50, §G.1 · p. 50

G.4 Case 1: Setting a reminder.

  • In Case 1, humans shared attribute readings—A was more explicit about its limitations and B more direct—yet neither group agreed on the overall winner.Human pooled attribute agreement was 1.000.Source: §G.4 · p. 51

G.5 Case 2: Layering clothes; G.6 Case 3: A home workout routine; G.7 Case 4: Becoming a morning person.

  • Case 2 had unanimous overall preference for A despite differing human clarity judgments; in Case 3, all humans saw B as more specific and example-rich, while six saw A as more qualified.In Case 4, both groups identified B as more specific, but humans disagreed more about insightfulness.Source: §3 · p. 52, §G.6 · p. 53, §G.7 · p. 54

G.8 Case 5: Ideas for leftover bananas.

  • Humans unanimously judged A broader and B stronger in storytelling, but overall choices split 3 to 2 with two abstentions, so agreement on attributes did not determine a unique winner.Across 87 attributes, 26 had more A than B votes, 49 more B than A, and 12 had zero net signal; mean commitment was 0.36.Source: §G.8 · p. 55

G.9 Case 6: Care after a tooth extraction.

  • Both groups identified A as more specific and B as more safety-conservative, while their overall majorities differed; specificity does not establish that medical advice is correct.Human pooled attribute agreement was 0.857; across 87 attributes, 51 had more A than B votes, 27 more B than A, and 9 had zero net signal.Source: §G.9 · p. 56

G.10 Case 7: Choosing a stand mixer.

  • Humans agreed A was more concise and B stronger in storytelling and vividness, yet overall votes split 3 to 3 with one abstention.Human pooled attribute agreement was 1.000; across 87 attributes, 30 had more A than B votes, 49 more B than A, and 8 had zero net signal.Source: §G.10 · p. 57

G.11 Case 8: Arithmetic or a joke.

  • Both groups identified A as more humorous and B as more rigorous, but their overall majorities differed, consistent with different interpretations of whether the prompt called for playfulness or correction.Across 87 attributes, 23 had more A than B votes, 59 more B than A, and 5 had zero net signal; mean commitment was 0.59.Source: §G.11 · pp. 57–58

G.12 Case 9: Drafting a business agreement.

  • Humans favored B while the LLM majority favored A, despite agreement that A was warmer and more persuasive and B more direct; A was truncated in the source dataset.Across 87 attributes, 72 had more A than B votes, 13 more B than A, and 2 had zero net signal; mean commitment was 0.52.Source: §G.12 · p. 59

G.13 Case 10: Books on Inuit mythology.

  • Both groups tended to identify A as more specific and relevant and B as more insightful; questioning the premise did not make a response the preferred answer to the request.Human pooled attribute agreement was 0.810; across 87 attributes, 42 had more A than B votes, 39 more B than A, and 6 had zero net signal.Source: §G.13 · pp. 59–60

G.14 Case 11: Responding to a fictional character.

  • Every human who chose a winner selected B, while the LLM majority selected A, despite shared judgments about the role-play’s interest and creativity and A’s explicit limitations.Across 87 attributes, 30 had more A than B votes, 53 more B than A, and 4 had zero net signal; mean commitment was 0.43.Source: §G.14 · p. 60

G.15 Case 12: Explaining inflation.

  • Both groups identified A as more accessible and stronger in analogy use and B as more informative, but these response judgments do not independently verify the economic explanation.Across 87 attributes, 23 had more A than B votes, 56 more B than A, and 8 had zero net signal; mean commitment was 0.53.Source: §G.15 · p. 61

G.16 Case 13: A stolen-lamps pun.

  • Humans identify A as more humorous and B as more empathetic, yet the human majority prefers B while the LLM majority prefers A.Human pooled attribute agreement is 1.000; across 87 attributes, 22 favor A, 62 favor B, and 3 have zero net signal, with mean commitment 0.65.Source: §G.16 · p. 62

G.17 Case 14: Imagining a first painting.

  • Every human resolving an attribute identifies A as using more examples and B as more creative; all five humans with an overall preference choose B, while two abstain.Human pooled attribute agreement is 1.000; across 87 attributes, 37 favor A, 45 favor B, and 5 have zero net signal, with mean commitment 0.42.Source: §G.17 · p. 63

G.18 Case 15: Responding to movie plans.

  • Both groups see A as broader and more informative and B as warmer, but their overall majorities differ, consistent with different interpretations of the prompt.Across 87 attributes, 64 favor A, 17 favor B, and 6 have zero net signal; mean commitment is 0.51. The groups may treat the prompt as either an information request or a conversational opening.Source: §G.18 · pp. 63–64

G.19 Case 16: Making a sentence more impactful.

  • Both groups judge A more vivid and B more concise, while human persuasiveness judgments are mixed and the stronger-sounding rewrite may shift the original emphasis.Human pooled attribute agreement is 0.841; across 87 attributes, 74 favor A, 11 favor B, and 2 have zero net signal, with mean commitment 0.54.Source: §G.19 · p. 64

G.20 Case 17: Describing hobbies.

This section lists individual overall judgments (A, B, or missing) for 23 cases across H1–H7 and notes that humans can agree on attribute judgments while dividing over overall preference. In the addition-only control, humans unanimously prefer B, and 20 of 21 LLMs vote for B.

G.20 Case 17: Describing hobbies.

  • LLMs identify B as more humorous, but human humor votes split evenly, showing that hidden consensus is substantial but not universal.Human pooled attribute agreement is 0.583. Across 87 attributes, 60 have more A than B votes, 12 more B than A, and 15 zero net signal; mean commitment is 0.31.Source: §G.20 · p. 65

G.21 Case 18: Explaining a kilowatt-hour.

  • All humans identify A as more concise and B as using more examples, yet six prefer B overall while LLM votes are more divided.Across 87 attributes, 18 have more A than B votes, 54 more B than A, and 15 zero net signal; mean commitment is 0.31.Source: §G.21 · p. 66

G.22 Case 19: Reversing yes and no.

  • All humans identify A as more direct and B as more qualified, but their overall votes split over answers that accommodate the local rule to different degrees.Human pooled attribute agreement is 1.000; across 87 attributes, 28 have more A than B votes, 52 more B than A, and 7 zero net signal; mean commitment is 0.53.Source: §G.22 · p. 67

G.23 Case 20: Lessons from experience.

  • All humans identify A as more concise and B as warmer and better justified, while most prefer B overall and a minority still selects A.Human pooled attribute agreement is 1.000; across 87 attributes, 9 have more A than B votes, 73 more B than A, and 5 zero net signal; mean commitment is 0.48.Source: §G.23 · pp. 67–68

G.24 Case 21: Why pancakes brown faster.

  • All humans identify A as more concise and B as greater in explanatory depth, while overall preferences are less unanimous between the short and longer explanations.Human pooled attribute agreement is 1.000; across 87 attributes, 9 have more A than B votes, 76 more B than A, and 2 zero net signal; mean commitment is 0.58.Source: §G.24 · pp. 68–69

G.25 Case 22: Eight eights make 1,000.

  • The addition-only control has unanimous human preference for B and 20 of 21 LLM votes for B, because A uses multiplication and subtraction despite the instruction.The expanded LLM panel is nearly, but not fully, unanimous; human pooled attribute agreement is 0.833.Source: §G.25 · p. 69

G.26 Case 23: Completing a rhyme.

  • All humans identify A as more qualified and B as more direct, but their overall votes split, so recognizing caution and directness does not settle their relative value.Human pooled attribute agreement is 1.000; across 87 attributes, 43 have more A than B votes, 39 more B than A, and 5 zero net signal; mean commitment is 0.53.Source: §G.26 · pp. 70–71

G.27 Individual overall judgments.

  • The individual-judgment tables record 21 LLM judges’ overall choices across 23 cases, retaining the same judge identifiers across cases and anonymizing human identities.A and B retain original dataset coordinates, T marks an LLM tie, and – marks a human abstention; the seven human columns preserve within-rater patterns.Source: §G.27 · p. 72
  • The individual-judgment table records A, B, or missing entries for each of 23 numbered cases across H1–H7.The supplied passages do not define what A, B, or the H labels mean.Source: §G.27 · p. 72

Related papers

Loading 2609.36705v1…