PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Preference Bias in Rubric-Based Evaluation of Large Language Models

José Pombal, Ricardo Rei, André F. T. Martins

arXiv:2604.06996v3cs.CLcs.AI

TL;DR

Self-preference bias in LLM judges had not been systematically studied in rubric-based evaluation, despite its growing use. This study evaluates the bias with objective and subjective rubrics and finds it persists across both settings, with rubric characteristics associated with variation in susceptibility.

  • Problem

    Whether rubric-based evaluation is susceptible to self-preference bias had not been systematically studied, despite its growing use for benchmarking and reward modeling.

    Source: §1 · p. 2

  • Method

    The paper systematically studies self-preference bias in rubric-based evaluation using objective rubrics and subjective medical-evaluation rubrics.

    Source: §1 · p. 2

  • Results

    Self-preference bias persists with objective and subjective rubrics; ensembling judges helps but does not eliminate it, and negative or communication-related rubrics are more susceptible.

    Source: §1 · p. 2, §4 · p. 8

  • Takeaways & Limitations

    Negative-rubric and topic-level patterns, along with inter-judge agreement, offer guidance for designing rubrics more resilient to self-preference bias.

    Source: §1 · p. 2

    An ensemble reference may underestimate self-preference bias because it represents all model families.

    Source: §3 · p. 5

Abstract

from arXiv · show

LLM-as-a-judge has become the de facto approach for evaluating LLM outputs. However, judges are known to exhibit self-preference bias (SPB): they tend to favor outputs produced by themselves or by models from their own family. This skews evaluations and, thus, hinders model development, especially in settings of recursive self-improvement. We present the first study of SPB in rubric-based evaluation, an increasingly popular benchmarking paradigm where judges issue binary verdicts on individual evaluation criteria, instead of assigning holistic scores or rankings. Using IFEval and LiveCodeBench, benchmarks with programmatically verifiable rubrics, we show that SPB persists even when evaluation criteria are entirely objective: among rubrics where generators fail, judges can be more than 50% more likely to incorrectly mark them as satisfied when the output is their own. We also find that, similarly to other evaluation paradigms, ensembling multiple judges helps mitigate SPB, but without fully eliminating it. On HealthBench, a medical chat benchmark with subjective rubrics, we observe that SPB skews model scores by up to 10 points, a potentially decisive margin when ranking frontier models. We analyze the factors that drive SPB in this setting, finding that negative rubrics and subjective topics like communication and emergency referrals are particularly susceptible.

1 Introduction

Rubric-based evaluation is increasingly used, but whether it avoids self-preference bias had not been systematically studied. The study finds bias persists with objective rubrics and in subjective medical evaluation, with rubric-level patterns that may guide more resilient evaluation.

1 Introduction

  • Rubric-based evaluation uses binary criteria to assess outputs and offers more granular, interpretable evaluations than pairwise comparison or numeric ratings.It is increasingly used for benchmarking and reward modeling, and has been reported to correlate well with human judgments.Source: §1 · p. 1
  • Unlike pairwise comparison and direct assessment, rubric-based evaluation avoids output ordering and score calibration issues, but its susceptibility to self-preference bias had not been systematically studied.This motivated examining whether rubric-based evaluation might be less vulnerable to bias.Source: §1 · p. 2
  • For failed rubrics, judges can be more than 50% more likely to incorrectly mark their own or family models’ outputs as satisfied, even with objective criteria and controls for difficulty and judge quality.The study also finds rubric-based bias is similar to direct assessment, less severe than pairwise comparison, and reduced—but not eliminated—by ensembling judges.Source: §1 · p. 2
  • On HealthBench, GPT-5 overestimates its performance by about 4 points relative to an ensemble, a potentially substantial margin when ranking frontier models.The study reports similar conclusions for this benchmark’s subjective, human-written rubrics.Source: §1 · p. 2
  • Negative rubrics and rubrics about communication or emergency referrals show higher self-preference bias; inter-judge agreement can help flag high-bias rubrics.These patterns offer guidance for designing rubrics more resilient to bias.Source: §1 · p. 2

2 Experimental setup

The section defines self-preference through overestimation relative to unrelated generators, then describes its accuracy measures, uncertainty estimation, model coverage, and benchmark datasets.

2.1 Measuring self-preference bias

  • HSPP Ratio compares a judge’s overestimation of its own outputs with its overestimation of unrelated generators; 1 indicates no self-preference and values above 1 indicate it.The family variant measures overestimation of same-family generators excluding the judge itself.Source: §2.1 · p. 3

2.2 Measuring performance

  • Performance is measured by instance-level pairwise concordance across generator pairs and, for single- and all-rubric evaluation, the fraction of rubric verdicts matching the reference.Pairwise comparisons are run in both presentation orders to mitigate position bias.Source: §2.2 · p. 4

2.3 Statistical significance

  • All metrics receive 95% bootstrap confidence intervals from 1,000 resamples, with the resampling unit matched to the metric’s granularity.Paired comparisons share resamples so the interval for their difference is valid.Source: §2.3 · p. 4

2.4 Models

  • The evaluation uses open- and closed-weight judges and generators spanning Gemma 3, Llama 4, Qwen 3, GPT-5, GPT-oss-120B, and Claude 4.5.The model set includes multiple sizes for Gemma 3 and Qwen 3, and Maverick and Scout variants of Llama 4.Source: §2.4 · p. 5

2.5 Datasets

  • IFEval and LiveCodeBench provide programmatically verifiable rubrics for objective-setting analyses, while HealthBench supplies subjective medical-chat rubrics.HealthBench has 5,000 instances and 48,562 rubrics; because its ground truth is unavailable, the study uses majority votes from the largest judge of each family as reference.Source: §2.5 · p. 5

3 Self-preference bias with objective rubrics

Rubric-based evaluation exhibits self-preference bias even with objective criteria, and its severity varies across evaluation paradigms. Ensembling and filtering low-agreement rubrics can mitigate the bias, but do not fully eliminate it.

3 Self-preference bias with objective rubrics

  • On LiveCodeBench, GPT-5 can be 20 times more likely to overestimate its own outputs; on IFEval, all but one judge has a mean self-preference ratio above 1.On IFEval, statistical significance holds for only two systems.Source: §3 · p. 5, Table 1 · p. 6
  • Self-preference bias is significantly more severe in pairwise comparison than rubric-based evaluation, while rubric-based evaluation is not necessarily more robust than direct assessment.Single-rubric presentation elicits slightly higher bias than presenting all rubrics at once, possibly because the narrower scope makes failures easier to rationalize away.Source: §3 · pp. 5–6, Figure 2 · p. 6
  • Ensembling yields the same or higher accuracy and lower self-preference bias, but does not eliminate it; on LiveCodeBench, reductions are significant for four of five committee members.On IFEval, reductions are mostly negative but not individually significant.Source: §3 · pp. 6–7
  • Removing rubrics with low inter-judge agreement can reduce self-preference bias, making agreement a practical signal for flagging problematic rubrics.Figure 4 compares instance- and rubric-level bias as the agreement threshold changes.Source: §3 · p. 7, Figure 4 · p. 7
  • After stratifying LiveCodeBench generations by difficulty, self-preference remains high across bins for all but Qwen 235B, indicating it persists beyond generator capability.Difficulty is proxied by mean rubric accuracy across the 12 judges.Source: §3 · p. 7, Table 2 · p. 8

4 Self-preference bias with subjective rubrics

HealthBench results show that self- and family-preference can materially shift scores, and that overestimation varies across rubric properties and topics.

4 Self-preference bias with subjective rubrics

  • Nearly 11 points: Gemma judges particularly overestimate Gemma-4B on HealthBench, while GPT-5 gives itself an approximately 4-point relative bonus.Llama judges show no strong self- or family-preference there, consistent with below-1 HSPP ratios.Source: §4 · p. 8
  • Negative rubrics increase self-overestimation, while communication, global health, and emergency-referral themes amplify it more; instruction-following elicits comparatively low bias.The regression models rubric-level overestimation on the error-eligible population, with judge and instance random intercepts; Figure 6 plots self × property interaction coefficients.Source: §4 · p. 9, Figure 6 · p. 9

5 Related work

Rubric-based evaluation offers granular binary judgments, but its relationship to self-preference bias had not been systematically studied. Prior work examined self-preference in pairwise comparison and direct assessment; this paper extends the analysis to rubric-based evaluation.

5.1 Rubric-based evaluation

5.2 Self-preference bias

  • Prior SPB studies examined pairwise comparison and direct assessment; this work finds the bias persists in rubric-based evaluation, even with objective rubrics, while judge ensembling only partly mitigates it.Other cited work links SPB to self-recognition, examines objectively worse self-outputs (Chen et al., 2025), and proposes jury evaluation (Verga et al., 2024).Source: §5.2 · p. 10

6 Conclusion

The study finds that rubric-based evaluation remains vulnerable to self-preference, even with objective rubrics, and that the bias can skew scores in subjective settings. The conclusion calls for mitigation beyond rubric construction, while the appendix reports family-level results.

6 Conclusion

  • Rubric-based evaluation is more robust than pairwise comparison but not immune to self-preference; ensembling helps without eliminating it, and fuller mitigation may require inference-time or training-time interventions.The conclusion identifies reward-modeling pipelines as a natural next step for analysis.Source: §6 · p. 10

D Family-level self-preference

  • On LiveCodeBench, GPT-5 has a family-level HSPP ratio of 11.91, and most judges over-credit same-family generators.The reported family-level ratios exclude the judge itself; their overall pattern mirrors the self-level results.Source: D Family-level self-preference · p. 17, Table 5 · p. 18

E Instance-level results on IFEval

The IFEval instance-level results are reported per judge, with confidence intervals.

E Instance-level results on IFEval

  • Table 6 reports IFEval single-rubric instance-level metrics per judge, with 95% confidence intervals computed using 1,000 instance-level bootstrap resamples.The table covers mean instance pairwise accuracy and instance-level HSPP ratios for self and family.Source: Table 6 · p. 18, E Instance-level results on IFEval · p. 18

F Ensembling: committee impact and composition

Majority-vote committees reduce self-preference for most LiveCodeBench members, while IFEval reductions are mostly not individually significant. Judge quality is more influential than family diversity, and correlated judge errors help explain the bias that remains.

F Ensembling: committee impact and composition

  • On LiveCodeBench, ensembling significantly reduces self-preference for four of five committee members; IFEval reductions are mostly negative but not individually significant.Tables 7 and 8 compare each member's individual HSPP-Self with the committee's HSPP on that member's outputs against the same non-family comparator set.Source: F Ensembling: committee impact and… · p. 18, Table 8 · p. 19, Table 7 · p. 19
  • Judge quality dominates committee composition on both benchmarks, while family diversity has a small positive association on IFEval and no significant association on LiveCodeBench.The analyses regress committee-mean HSPP-Rub. Self on committee size, family count, and mean member MRA.Source: F Ensembling: committee impact and… · p. 19, Table 9 · p. 19
  • Even where all generators fail, judges agree above chance (α = 0.308 on IFEval and α = 0.154), reflecting shared false positives that help explain bias surviving majority voting.Because ground truth is constant in this subset, positive agreement means judges mark the same failed rubrics as passed.Source: F Ensembling: committee impact and… · p. 19

G Judge reasoning-effort ablation

Increasing reasoning effort slightly improves MRA but does not reduce self-preference on IFEval. Capability analyses find that self-preference attenuates in some comparisons but never disappears, while stronger generators' failures tend to be harder to judge.

G Judge reasoning-effort ablation

  • On IFEval, increased reasoning effort slightly improves MRA but does not reduce self-preference; if anything, self-preference increases.The ablation varies reasoning effort for two proprietary judges.Source: G Judge reasoning-effort ablation · p. 19

H Generator-capability analyses

  • Across three complementary generator-capability analyses, self-preference attenuates in places but never disappears.The analyses include same-quality-tertile comparators and GPT-5 comparisons against similar-capability generators, including rubrics both generators objectively fail.Source: H Generator-capability analyses · p. 20
  • On LiveCodeBench, stronger generators' failed completions concentrate in harder cells, while weaker generators' failures concentrate in easier cells.The full hardness-stratified analysis extends the approach to all twelve judges and all three benchmarks, binning by per-generation mean MRA.Source: H Generator-capability analyses · p. 20

I Output-property mechanisms

Output length, familiarity, and self-recognition are tested as possible contributors to self-preference. Length and recognition track differences in bias, while controlling for familiarity does not eliminate it in the lowest-perplexity quartile.

I Output-property mechanisms

  • Self-preference grows with output length on HealthBench and LiveCodeBench.The analysis compares aggregate HSPP-Rub. Self across within-judge output-length quartiles for all twelve judges.Source: I Output-property mechanisms · p. 20
  • In the lowest-perplexity quartile, self-preference remains above 1 on LiveCodeBench (1.48) and HealthBench (1.03); upper-quartile estimates are unreliable.The upper quartiles are unreliable because judges' own outputs rarely fall there.Source: I Output-property mechanisms · p. 21
  • On LiveCodeBench, self-preference is 1.88 when judges recognize their own outputs and 0.72 when they do not.The analysis splits judges' own outputs by their prediction of whether each output was self-generated.Source: I Output-property mechanisms · p. 21

J HealthBench reference ablations

HealthBench's fixed five-family reference is lenient because it includes every judge's family, making measured self-preference conservative. The section also examines regression interactions, the counts behind HSPP ratios, and confidence intervals for centered score deltas.

J HealthBench reference ablations

  • For GPT-oss-120B, HealthBench HSPP-Rub. Self rises from 1.00 to 1.26/1.30 under LOFO/LOO references; for GPT-5, it rises from 1.54 to 1.67/1.68.Because every judge's family appears in the fixed reference, the paper characterizes its self-preference estimates as conservative.Source: J HealthBench reference ablations · p. 21, Table 23 · p. 25

K Mixed-effects regression details

  • Regression interactions show stronger self-specific overestimation for negative rubrics and the Communication, Emergency Referrals, Global Health, and Complex Responses themes; length and rubric axes show no self-specific effect.The pooled self main effect is indistinguishable from zero, as self-preferring and self-deprecating judges cancel out and the ensemble reference partly absorbs judge leniency.Source: K Mixed-effects regression details · p. 22, Table 24 · p. 25

L Absolute values behind the HSPP ratios

  • HSPP ratio tables expose own-output false-positive and eligible-rubric counts in the numerator, alongside pooled non-family comparator counts and the mean per-generator comparator false-positive rate as denominator..Source: L Absolute values behind the HSPP ratios · p. 22

M Significance of the HealthBench score delta matrix

  • For the centered HealthBench score-delta matrix, all diagonal self-evaluation cells are significant except Llama-Mav's and Qwen-235B's.Table 28 reports instance-bootstrap 95% confidence intervals; cells whose intervals include zero are shown in gray.Source: M Significance of the HealthBench… · p. 22, Table 28 · p. 27

Related papers

Loading 2604.06996v3…