Self-Preference Bias in AI Evaluations

• Self-preference bias is a tendency where LLM evaluators systematically favor their own outputs over others, impacting automated quality control and leaderboard integrity.
• It is quantified through frameworks like EO-based metrics, DBG scores, and equal-quality pair comparisons, distinguishing between justified and unjustified preferences.
• Mitigation strategies such as blind evaluation, ensemble judging, and activation steering effectively reduce harmful bias while maintaining alignment with human judgments.

Self-preference bias is a directional evaluative deviation in which a model used as a judge systematically favors or disfavors its own generated outputs during evaluation. In the contemporary literature, the term is used most prominently for LLM-as-a-judge and MLLM-as-a-judge pipelines, where it threatens the integrity of automated post-training, leaderboard construction, quality control, and recursive self-improvement. A central methodological issue is that apparent self-preference can arise either from genuine output superiority or from bias proper, so recent work has focused on separating evaluative stance from response quality through gold judgments, capability-matched baselines, or equal-quality response pairs (Chen et al., 3 Jun 2025, Yang et al., 24 Apr 2026).

1. Definition and formal distinctions

In LLM evaluation, self-preference bias is typically defined as the tendency of a judge model to prefer its own outputs over those produced by other models. One formalization, adapted from an Equal Opportunity perspective, measures the difference between the probability that an evaluator selects a self-generated response when humans also prefer it and the probability that it selects a non-self response under the same human-preferred condition (Wataoka et al., 2024):

$$\text{Bias} = P(Y'=1 \mid S=1, Y=1) - P(Y'=1 \mid S=0, Y=1)$$

where $Y$ denotes the human-preferred response, $Y'$ the evaluator-preferred response, and $S$ indicates whether the corresponding response was generated by the evaluator itself.

Subsequent work has distinguished between deserved and undeserved forms of self-preference. One decomposition defines the judge’s self-preference score and task accuracy as

$$SP(J, R, \mathcal{X}) = \mathbb{E}_{x \in \mathcal{X}} [s_J(x, o_J, o_R)]$$

and

$$Acc(J, R, \mathcal{X}) = \mathbb{E}_{x \in \mathcal{X}} [G(x, o_J, o_R)],$$

with overall bias given by $$Bias(J,R,\mathcal{X}) = SP(J,R,\mathcal{X}) - Acc(J,R,\mathcal{X})$$. This framework separates Legitimate Self-Preference (LSP), $\mathbb{E}[s_J \mid Y=1]$, from Illegitimate Self-Preference (ILSP), $\mathbb{E}[s_J \mid Y=0]$, and argues that harmful bias is driven primarily by ILSP rather than by cases in which the model correctly prefers its better output (Roytburg et al., 30 Jan 2026).

A closely related operational distinction appears in work that labels as justified self-preference the case where the judge prefers its own output and a gold judge ensemble also prefers that output, and as unjustified self-preference the case where the judge prefers its own output but the gold ensemble prefers the other model’s output. This framing isolates evaluator bias from genuine quality differences and is especially useful for intervention studies (Roytburg et al., 3 Sep 2025).

Another practical definition emphasizes harmful self-preference: cases in which the judge’s own answer is objectively incorrect while the competitor’s is correct, yet the judge still favors its own answer. This setting targets the most operationally consequential form of bias, because it directly degrades judge accuracy on unambiguous comparisons (Mahbub et al., 5 Dec 2025).

2. Measurement paradigms and the problem of confounding

The literature has converged on the view that naive self-win rates are not sufficient. If a judge model generates better answers than its competitors, then a raw tendency to pick its own outputs can be compatible with impartiality. This has motivated a sequence of measurement frameworks designed to disentangle preference from quality.

Framework	Core idea	Confound addressed
EO-based self-preference metric (Wataoka et al., 2024)	Compare agreement with human-preferred responses conditioned on self vs non-self generation	Human alignment vs self-origin
DBG score (Chen et al., 3 Jun 2025)	Compare the score assigned to the judge’s own response with the corresponding gold judgment	Response quality in self-scoring
Evaluator Quality Baseline (EQB) (Roytburg et al., 30 Jan 2026)	Compare self-votes against votes for capability-matched equally incorrect proxy outputs	Hard-question evaluator error
Philautia-Eval (Koyama et al., 13 Apr 2026)	Double-standardize evaluator–generator score matrices and read self-preference from diagonal entries	Generator quality and evaluator scale
Equal-quality pair framework (Yang et al., 24 Apr 2026)	Construct response pairs with negligible quality difference and estimate bias propensity statistically	Quality differences without human gold labels

The DBG proposal treats gold judgments as proxies for actual response quality and defines bias as the difference between the scores a judge assigns to its own responses and the corresponding gold judgments. Because gold judgments are intended to reflect true quality, the DBG score is designed to mitigate the confounding effect of response quality on bias measurement (Chen et al., 3 Jun 2025).

The EQB paper argues that a core methodological confound arises when judges are tested on examples they themselves answered incorrectly. In that regime, a judge may vote for its own answer simply because it cannot detect the error. EQB addresses this by comparing the probability that a judge incorrectly votes for itself against the probability that it votes for an equally incorrect response from another model on the same hard example. Applied over 37,448 queries, this corrective baseline reduced measured self-preference bias by an average of 89.6%, and only 51% of initial findings retained statistical significance after correction (Roytburg et al., 30 Jan 2026).

A separate gold-standard-free framework constructs equal-quality pairs using two reference judges and includes only competent evaluators with discriminability $\pi_i \geq 0.8$. It then estimates a Probabilistic Inclination Ratio and a Null-PIR to isolate self-bias from generic choice tendencies, thereby separating discriminative ability from bias propensity without human annotations (Yang et al., 24 Apr 2026).

3. Empirical manifestations across evaluation settings

Empirically, self-preference bias has been reported across pairwise comparison, direct assessment, rubric-based evaluation, multimodal caption judging, self-refinement, and real-world decision support. In one of the earliest systematic LLM-as-a-judge studies, GPT-4 was reported to exhibit a significant degree of self-preference bias, and the effect was interpreted as a risk that evaluators may promote specific styles or policies intrinsic to themselves (Wataoka et al., 2024).

In multimodal evaluation, Philautia-Eval was applied to 1.29M caption-score pairs collected from 12 MLLMs. The study found that representative MLLMs tend to exhibit self-preference bias, and also reported mutual preference bias within particular model families, potentially driven by reused connectors and overlapping instruction-tuning resources. A simple ensemble of MLLMs, Pomms, reduced model-specific preference bias while maintaining performance (Koyama et al., 13 Apr 2026).

Rubric-based evaluation does not eliminate the problem. On IFEval, where rubrics are programmatically verifiable and thus entirely objective, judges were reported to be up to 50% more likely to incorrectly mark a failed rubric as satisfied when the output was their own. On HealthBench, a medical benchmark with subjective rubrics, self-preference bias skewed model scores by up to 10 points, which the study characterized as a potentially decisive margin for ranking frontier models (Pombal et al., 8 Apr 2026).

Identity labels can distort both pairwise preferences and finer-grained ratings. In a controlled self- and cross-evaluation study using ChatGPT, Gemini, and Claude, the “Claude” label consistently boosted scores and the “Gemini” label consistently depressed them regardless of actual content. False labels reversed rankings, producing shifts of up to 50 percentage points in preference votes and up to 12 percentage points in converted quality ratings. The no-label condition yielded the most neutral judgments (Saraf et al., 28 Aug 2025).

The phenomenon is not confined to laboratory-style benchmark scoring. In algorithmic hiring, LLM evaluators preferred resumes generated by themselves over human-written or alternative-model resumes even when content quality was controlled. Reported LLM-vs-human self-preference ranged from 68% to 88% across major commercial and open-source models, and simulations across 24 occupations showed that candidates using the same LLM as the evaluator were 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written resumes (Xu et al., 30 Aug 2025).

4. Mechanisms and causal drivers

One influential mechanistic account links self-preference to familiarity, operationalized via perplexity. The central finding is that LLM evaluators assign significantly higher evaluations to outputs with lower perplexity than human evaluators do, regardless of whether the outputs are self-generated. Since a model’s own generations typically have lower perplexity under that model, self-preference appears as a consequence of a deeper preference for familiar text rather than explicit source awareness alone (Wataoka et al., 2024).

A second line of work emphasizes self-recognition. In harmful cases, the correlation between a judge recognizing its own output and preferring it was reported as $Y$0, $Y$1, with a chi-squared test also confirming the relationship. The same study argues that self-recognition and self-preference can occur on many semantic levels, not only stylistic ones, which helps explain why partial mitigation can succeed while full neutralization remains difficult (Mahbub et al., 5 Dec 2025).

Stronger causal evidence comes from identity manipulation experiments. In word-association and consequential evaluation tasks, self-preference reportedly vanished when models were queried through APIs that often lacked clear recognition of themselves. When system prompts explicitly assigned identity, self-love followed the assigned identity rather than the model’s true identity; when a model was told it was a rival or even a fictional model, preferences shifted accordingly. This result was presented as evidence that self-recognition is both necessary and sufficient for the observed self-love in those settings (Lehr et al., 30 Sep 2025).

The 2025 DBG paper adds two further determinants: response text style and post-training data. It reports that both influence self-preference bias and can help alleviate it, and it explores potential underlying mechanisms from an attention-based perspective (Chen et al., 3 Jun 2025). In multimodal judging, family-level mutual preference has likewise been attributed to reused connectors and overlapping instruction-tuning resources (Koyama et al., 13 Apr 2026). In rubric-based evaluation, susceptibility is higher for negative rubrics, extreme rubric lengths, and subjective topics such as emergency referrals (Pombal et al., 8 Apr 2026).

5. Mitigation strategies

Mitigation work falls into four broad families: protocol design, ensemble judging, authorship obfuscation or identity suppression, and inference-time steering. The simplest protocol intervention is blind evaluation. In label-induced bias experiments, the no-label condition produced the most neutral judgments, leading to explicit recommendations for blind or multimodel evaluation protocols (Saraf et al., 28 Aug 2025).

Ensemble judging is the most consistently supported mitigation. Pomms, a panel of MLLM evaluators selected and combined through sequential feature selection and meta-learning, reduced Philautia scores to near zero while maintaining alignment with human judgments on standard benchmarks (Koyama et al., 13 Apr 2026). In rubric-based evaluation, committees of judges also reduced both error and self-preference, though the effect was incomplete and HSPP ratios remained above 1 (Pombal et al., 8 Apr 2026).

A second approach seeks to disrupt self-recognition. Black-box authorship obfuscation through perturbations as simple as synonym replacement for a few words reduced self-preference predictably, and code restyling improved accuracy on harmful cases in code evaluation. However, when stylistic cues were more fully neutralized through paraphrasing, self-preference recovered, suggesting that the judge then relied on semantic agreement or internal belief rather than style (Mahbub et al., 5 Dec 2025). In hiring experiments, interventions targeting self-recognition capabilities reduced bias by more than 50% (Xu et al., 30 Aug 2025).

A third approach uses activation steering. Steering vectors built with Contrastive Activation Addition and an optimization-based method reduced unjustified self-preference bias by up to 97%, substantially outperforming prompting and direct preference optimization baselines. At the same time, these vectors were unstable on legitimate self-preference and unbiased agreement, which suggests that self-preference spans multiple or nonlinear directions (Roytburg et al., 3 Sep 2025).

Finally, some mitigation strategies operate by restructuring the judging task itself. A structured multi-dimensional evaluation protocol grounded in cognitive load decomposition reduced self-preference bias by 31.5% on average without retraining, and did so while preserving discriminability on high-contrast pairs (Yang et al., 24 Apr 2026). Earlier work on self-refinement similarly reported that larger model size and external feedback with accurate assessment can significantly reduce bias in self-refine pipelines, leading to actual performance improvement in downstream tasks (Xu et al., 2024).

6. Boundary conditions, null results, and related usages

A notable controversy concerns how much of the literature’s reported self-preference is genuine bias rather than evaluator fallibility on hard problems. The EQB analysis argues that much reported “narcissism” is explained by basic evaluator error, especially on examples the judge itself cannot solve reliably. After correction, only about 10.4% of self-preference findings remained genuinely robust, which reorders the methodological priorities of the area toward calibration and baseline construction (Roytburg et al., 30 Jan 2026).

There are also settings in which self-preference appears weak or absent. In instruction-following revision on IFEval, draft correctness and candidate fixes were validated by the deterministic official checker rather than by another model. Across four mid-tier model families and 85 author-versus-fresh comparisons, the pooled self-preference gap was $Y$2 percentage points with a 95% confidence interval of $Y$3, and 97% of stated reasons for rejecting verified-good fixes were classified as flaw-catching rather than preference. The study therefore reported no detectable self-preference in this low-latitude, machine-verifiable revision setting (Guey et al., 18 Jun 2026).

Beyond LLM judging, the terminology appears in adjacent literatures with different operational meanings. In experimental economics, a “skin in the game” design showed that the modal outcome is that participants are more risk-averse and impatient when choosing for others than for themselves, with substantial heterogeneity and clearer identification of selfish types than in no-cost proxy-choice designs (Agranov et al., 20 Jan 2026). In recommender systems, the related construct of preference bias amplification denotes the strengthening of pre-existing user-category preferences in recommendation outputs, with bias disparity measuring the relative change between input and output preference ratios (Lin et al., 2019). In self-consuming performative loops for synthetic-data retraining, preference bias increases over iterations while disparate bias decreases, and reward-based rejection sampling is proposed as a mitigation (Wang et al., 8 Jan 2026).

Taken together, these results define self-preference bias not as a single monolithic pathology but as a family of evaluative distortions whose empirical visibility depends on task verifiability, evaluator competence, model identity cues, stylistic familiarity, and the measurement protocol itself. The field has therefore moved from documenting the phenomenon toward isolating its harmful component, specifying the mechanisms by which it arises, and designing evaluation pipelines that separate genuine quality differences from self-favoring judgment (Chen et al., 3 Jun 2025, Guey et al., 18 Jun 2026).