Bias in LLM-as-a-Judge

Background: Accurate translation of radiology reports is important for multilingual research, clinical communication, and radiology education, but the validity of LLM-based evaluation remains unclear. Objective: To evaluate the educational suitability of LLM-generated Japanese translations of chest CT reports and compare radiologist assessments with LLM-as-a-judge evaluations. Methods: We analyzed 150 chest CT reports from the CT-RATE-JPN validation set. For each English report, a human-edited Japanese translation was compared with an LLM-generated translation by DeepSeek-V3.2. A board-certified radiologist and a radiology resident independently performed blinded pairwise evaluations across 4 criteria: terminology accuracy, readability, overall quality, and radiologist-style authenticity. In parallel, 3 LLM judges (DeepSeek-V3.2, Mistral Large 3, and GPT-5) evaluated the same pairs. Agreement was assessed using QWK and percentage agreement. Results: Agreement between radiologists and LLM judges was near zero (QWK=-0.04 to 0.15). Agreement between the 2 radiologists was also poor (QWK=0.01 to 0.06). Radiologist 1 rated terminology as equivalent in 59% of cases and favored the LLM translation for readability (51%) and overall quality (51%). Radiologist 2 rated readability as equivalent in 75% of cases and favored the human-edited translation for overall quality (40% vs 21%). All 3 LLM judges strongly favored the LLM translation across all criteria (70%-99%) and rated it as more radiologist-like in >93% of cases. Conclusions: LLM-generated translations were often judged natural and fluent, but the 2 radiologists differed substantially. LLM-as-a-judge showed strong preference for LLM output and negligible agreement with radiologists. For educational use of translated radiology reports, automated LLM-based evaluation alone is insufficient; expert radiologist review remains important.

As LLMs are increasingly used as judges in code applications, they should be evaluated in realistic interactive settings that capture partial context and ambiguous intent. We present TRACE (Tool for Rubric Analysis in Code Evaluation), a framework that evaluates LLM judges' ability to predict human preferences and automatically extracts rubric items to reveal systematic biases in how humans and models weigh each item. Across three modalities -- chat-based programming, IDE autocompletion, and instructed code editing -- we use TRACE to measure how well LLM judges align with developer preferences. Among 13 different models, the best judges underperform human annotators by 12-23%. TRACE identifies 35 significant sources of misalignment between humans and judges across interaction modalities, the majority of which correspond to existing software engineering code quality criteria. For example, in chat-based coding, judges are biased towards longer code explanations while humans prefer shorter ones. We find significant misalignment on the majority of existing code quality dimensions, showing alignment gaps between LLM judges and human preference in realistic coding applications.

Hallucination remains a critical bottleneck for large language models (LLMs), undermining their reliability in real-world applications, especially in Retrieval-Augmented Generation (RAG) systems. While existing hallucination detection methods employ LLM-as-a-judge to verify LLM outputs against retrieved evidence, they suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. To address this, we introduce Multi-Agent Reinforced Self-Check for Hallucination (MARCH), a framework that enforces rigorous factual alignment by leveraging deliberate information asymmetry. MARCH orchestrates a collaborative pipeline of three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates an initial RAG response, which the Proposer decomposes into claim-level verifiable atomic propositions. Crucially, the Checker validates these propositions against retrieved evidence in isolation, deprived of the Solver's original output. This well-crafted information asymmetry scheme breaks the cycle of self-confirmation bias. By training this pipeline with multi-agent reinforcement learning (MARL), we enable the agents to co-evolve and optimize factual adherence. Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucination rates. Notably, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models. MARCH paves a scalable path for factual self-improvement of LLMs through co-evolution. The code is at https://github.com/Qwen-Applications/MARCH.

Large language models (LLMs) are now widely used as judges, yet their decisions can change under presentation choices that should be irrelevant. We study one such source of instability: candidate-order sensitivity in listwise factuality evaluation, where several answers can look similarly polished while differing substantially in hallucination risk. We introduce PCFJudge, an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus decision. On RewardBench 2 Factuality, the final seven-permutation aggregate (K=7) improves top-1 selection accuracy from 86.00% to 91.33% with GPT-5.4 and from 86.33% to 89.67% with Claude Sonnet 4.6. These results suggest that candidate order can be a meaningful source of factuality-judging error and that marginalizing over this nuisance variation can improve the reliability of LLM evaluation.

As large language models (LLMs) are increasingly deployed as automated graders in educational settings, concerns about fairness and bias in their evaluations have become critical. This study investigates whether LLMs exhibit implicit grading bias based on writing style when the underlying content correctness remains constant. We constructed a controlled dataset of 180 student responses across three subjects (Mathematics, Programming, and Essay/Writing), each with three surface-level perturbation types: grammar errors, informal language, and non-native phrasing. Two state-of-the-art open-source LLMs -- LLaMA 3.3 70B (Meta) and Qwen 2.5 72B (Alibaba) -- were prompted to grade responses on a 1-10 scale with explicit instructions to evaluate content correctness only and to disregard writing style. Our results reveal statistically significant grading bias in Essay/Writing tasks across both models and all perturbation types (p < 0.05), with effect sizes ranging from medium (Cohen's d = 0.64) to very large (d = 4.25). Informal language received the heaviest penalty, with LLaMA deducting an average of 1.90 points and Qwen deducting 1.20 points on a 10-point scale -- penalties comparable to the difference between a B+ and C+ letter grade. Non-native phrasing was penalized 1.35 and 0.90 points respectively. In sharp contrast, Mathematics and Programming tasks showed minimal bias, with most conditions failing to reach statistical significance. These findings demonstrate that LLM grading bias is subject-dependent, style-sensitive, and persists despite explicit counter-bias instructions in the grading prompt. We discuss implications for equitable deployment of LLM-based grading systems and recommend bias auditing protocols before institutional adoption.

It is known that large language models (LLMs) underperform in English dialects, and that improving them is difficult due to data scarcity. In this work we investigate how quality and availability impact the feasibility of improving LLMs in this context. For this, we evaluate three rarely-studied English dialects (Yorkshire, Geordie, and Cornish), plus African-American Vernacular English, and West Frisian as control. We find that human-human agreement when determining LLM generation quality directly impacts LLM-as-a-judge performance. That is, LLM-human agreement mimics the human-human agreement pattern, and so do metrics such as accuracy. It is an issue because LLM-human agreement measures an LLM's alignment with the human consensus; and hence raises questions about the feasibility of improving LLM performance in locales where low populations induce low agreement. We also note that fine-tuning does not eradicate, and might amplify, this pattern in English dialects. But also find encouraging signals, such as some LLMs' ability to generate high-quality data, thus enabling scalability. We argue that data must be carefully evaluated to ensure fair and inclusive LLM improvement; and, in the presence of scarcity, new tools are needed to handle the pattern found.

Evaluating AI-generated research ideas typically relies on LLM judges or human panels -- both subjective and disconnected from actual research impact. We introduce HindSight, a time-split evaluation framework that measures idea quality by matching generated ideas against real future publications and scoring them by citation impact and venue acceptance. Using a temporal cutoff~$T$, we restrict an idea generation system to pre-$T$ literature, then evaluate its outputs against papers published in the subsequent 30 months. Experiments across 10 AI/ML research topics reveal a striking disconnect: LLM-as-Judge finds no significant difference between retrieval-augmented and vanilla idea generation ($p{=}0.584$), while HindSight shows the retrieval-augmented system produces 2.5$\times$ higher-scoring ideas ($p{<}0.001$). Moreover, HindSight scores are negatively correlated with LLM-judged novelty ($ρ{=}{-}0.29$, $p{<}0.01$), suggesting that LLMs systematically overvalue novel-sounding ideas that never materialize in real research.

The paradigm of LLM-as-a-judge relies on a critical assumption, namely that high inter-evaluator agreement indicates reliable and objective evaluation. We present two complementary findings that challenge this assumption. First, we demonstrate that this consensus is frequently illusory. We identify and formalize Evaluation Illusion, a phenomenon where LLM judges generate sophisticated critiques yet anchor scores on shared surface heuristics rather than substantive quality. Through a large-scale study of 105,600 evaluation instances (32 LLMs $\times$ 3 frontier judges $\times$ 100 tasks $\times$ 11 temperatures), we show that model-level agreement (Spearman $ρ= 0.99$) masks fragile sample-level agreement (Pearson $\bar{r} = 0.72$; absolute agreement ICC $= 0.67$), that merely sharing rubric structure restores 62% of total agreement, and that high-quality outputs paradoxically receive the least consistent evaluations. Second, we demonstrate that dynamically generating evaluation rubrics grounded in domain knowledge produces more meaningful assessment. We introduce MERG (Metacognitive Enhanced Rubric Generation), a knowledge-driven rubric generation framework whose domain-selective effects confirm this. Agreement increases in codified domains (Education +22%, Academic +27%) where knowledge anchors evaluators on shared standards, while it decreases in subjective domains where genuine evaluative pluralism emerges. These findings suggest that evaluation rubrics should be dynamically enriched with expert knowledge rather than relying on generic criteria, with implications for reward modeling in RLAIF.

Designing service systems requires selecting among alternative configurations -- choosing the best chatbot variant, the optimal routing policy, or the most effective quality control procedure. In many service systems, the primary evidence of performance quality is textual -- customer support transcripts, complaint narratives, compliance review reports -- rather than the scalar measurements assumed by classical optimization methods. Large language models (LLMs) can read such textual evidence and produce standardized quality scores, but these automated judges exhibit systematic biases that vary across alternatives and evaluation instances. Human expert review remains accurate but costly. We study how to identify the best service configuration with high confidence while minimizing expensive human audits, given that automated evaluation is cheap but biased. We formalize this as a sequential decision problem where a biased proxy score is observed for every evaluation, and a verified outcome can be acquired selectively at additional cost. We prove that LLM-only selection fails under arm-dependent bias, and that naive selective-audit estimators can be asymptotically biased. We develop an estimator combining proxy scores with inverse-propensity-weighted residuals and construct anytime-valid confidence sequences. Our algorithm, PP-LUCB, jointly decides which alternatives to evaluate and whether to request human audits, concentrating reviews where the LLM judge is least reliable. We prove correctness and establish instance-dependent cost bounds showing near-optimal efficiency. On a customer support ticket classification task, our algorithm correctly identifies the best model in 40/40 trials while achieving 90% audit cost reduction.

Large language models (LLMs) have become a standard for multilingual evaluation, yet they exhibit a severe systematic translationese bias. In this paper, translationese bias is characterized as LLMs systematically favoring machine-translated text over human-authored references, particularly in low-resource languages. We attribute this bias to spurious correlations with (i) latent manifold alignment with English and (ii) cross-lingual predictability. To mitigate this bias, we propose DIBJudge, a robust fine-tuning framework that learns a minimally sufficient, judgment-critical representation via variational information compression, while explicitly isolating spurious factors into the dedicated bias branch. Furthermore, we incorporate a cross-covariance penalty that explicitly suppresses statistical dependence between robust and bias representations, thereby encouraging effective disentanglement. Extensive evaluations on multilingual reward modeling benchmarks and a dedicated translationese bias evaluation suite demonstrate that the proposed DIBJudge consistently outperforms strong baselines and substantially mitigates translationese bias.

LLM-as-a-Judge is a flexible framework for text evaluation, which allows us to obtain scores for the quality of a given text from various perspectives by changing the prompt template. Two main challenges in using LLM-as-a-Judge are computational cost of inference using a large language model (LLM), especially when evaluating a large number of instances, and inherent bias of an LLM evaluator. To address these issues and reveal the structure of score bias caused by an LLM evaluator, we propose to apply a tensor clustering method to a given LLM-as-a-Judge score tensor, whose entries are the scores for different combinations of questions, answerers, and evaluators. Specifically, we develop a new tensor clustering method MultiwayPAM, with which we can simultaneously estimate the cluster membership and the medoids for each mode of a given data tensor. By observing the medoids obtained by MultiwayPAM, we can gain knowledge about the membership of each question/answerer/evaluator cluster. We experimentally show the effectiveness of MultiwayPAM by applying it to the score tensors for two practical datasets.

Reinforcement Learning from Human Feedback (RLHF) assumes annotator preferences reflect stable internal states. We challenge this through three experiments spanning the preference pipeline. In a human choice blindness study, 91% of surreptitiously swapped preferences go undetected, extending choice blindness to third-person evaluative comparison of unfamiliar text. Testing fifteen LLM judges as potential replacements, we find detection relies on shallow text matching rather than genuine self-monitoring: removing prior reasoning from context causes blindness to surge from near-zero to over 50%, while explicit social pressure induces near-universal compliance. In a dose-response experiment across two architectures from 86M to 2B parameters, one-sixth to one-third of labels must be corrupted before the reward signal halves, yet standard pairwise accuracy remains virtually unchanged. A Best-of-N evaluation confirms this translates to downstream policy degradation: at 50% corruption, reward-guided selection produces no improvement over random sampling, while the proxy model reports monotonically increasing scores. Together, these results reveal a preference construction problem: the signal entering RLHF is shaped by elicitation context in ways that neither human metacognition, LLM self-monitoring, nor standard evaluation metrics can detect.

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g., verbosity, position) in LLM evaluators, and creating a scalability-reliability trade-off. To address these limitations, we propose CDRRM (Contrast-Driven Rubric Reward Model), a framework built on a novel Contrast-then-Synthesis paradigm for high-quality rubric generation and guided preference judgment. CDRRM first conducts multi-dimensional contrastive profiling on preference pairs to identify causal discriminative factors, then synthesizes these insights into compact, context-aware rubrics to guide preference judg- ments. Extensive experiments on three authoritative benchmarks (RewardBench, RMBench, RMB) demonstrate that CDRRM achieves state-of-the-art performance across diverse domains and effectively mitigates aforementioned evaluation biases. Notably, our approach delivers exceptional data efficiency: training the rubric generator on only 3k high-quality samples empowers a frozen pre-trained judge model to outperform fully fine-tuned baselines. This work offers a scalable, interpretable, and data-efficient path for reward modeling.

Navigating complex social situations is an integral part of corporate life, ranging from giving critical feedback without hurting morale to rejecting requests without alienating teammates. Although large language models (LLMs) are permeating the workplace, it is unclear how well they can navigate these norms. To investigate this question, we created HR Simulator, a game where users roleplay as an HR officer and write emails to tackle challenging workplace scenarios, evaluated with GPT-4o as a judge based on scenario-specific rubrics. We analyze over 600 human and LLM emails and find systematic differences in style: LLM emails are more formal and empathetic. Furthermore, humans underperform LLMs (e.g., 23.5% vs. 48-54% scenario pass rate), but human emails rewritten by LLMs can outperform both, which indicates a hybrid advantage. On the evaluation side, judges can exhibit differences in their email preferences: an analysis of 10 judge models reveals evidence for emergent tact, where weaker models prefer direct, blunt communication but stronger models prefer more subtle messages. Judges also agree with each other more as they scale, which hints at a convergence toward shared communicative norms that may differ from humans'. Overall, our results suggest LLMs could substantially reshape communication in the workplace if they are widely adopted in professional correspondence.

As AI models progress beyond simple chatbots into more complex workflows, we draw ever closer to the event horizon beyond which AI systems will be utilized in autonomous, self-maintaining feedback loops. Any autonomous AI system will depend on automated, verifiable rewards and feedback; in settings where ground truth is sparse or non-deterministic, one practical source of such rewards is an LLM-as-a-Judge. Although LLM judges continue to improve, the literature has yet to introduce systems capable of enforcing standards with strong guarantees, particularly when bias vectors are unknown or adversarially discovered. To remedy this issue, we propose average bias-boundedness (A-BB), an algorithmic framework which formally guarantees reductions of harm/impact as a result of any measurable bias in an LLM judge. Evaluating on Arena-Hard-Auto with four LLM judges, we achieve (tau=0.5, delta=0.01) bias-bounded guarantees while retaining 61-99% correlation with original rankings across formatting and schematic bias settings, with most judge-bias combinations exceeding 80%. The code to reproduce our findings is available at https://github.com/penfever/bias-bounded-evaluation.

Automatic evaluation of medical open-ended question answering (OEQA) remains challenging due to the need for expert annotations. We evaluate whether large language models (LLMs) can act as judges of semantic equivalence in French medical OEQA, comparing closed-access, general-purpose, and biomedical domain-adapted models. Our results show that LLM-based judgments are strongly influenced by the model that generated the answer, with agreement varying substantially across generators. Domain-adapted and large general-purpose models achieve the highest alignment with expert annotations. We further show that lightweight adaptation of a compact model using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) substantially improves performance and reduces generator sensitivity, even with limited data. Overall, our findings highlight the need for generator-aware evaluation and suggest that carefully adapted small models can support scalable evaluation in low-resource medical settings.

LLM-as-judge evaluation has become standard practice for open-ended model assessment; however, judges exhibit systematic biases that cannot be averaged out by increasing the number of scenarios or generations. These biases are often similar in magnitude to the model differences that benchmarks are designed to detect, resulting in unreliable rankings when single-judge evaluations are used. We introduce a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, CyclicJudge, a round-robin assignment of judges to scenarios, is demonstrated to be the optimal strategy for a fixed judge panel and judge-call budget: the score recovers the panel mean exactly while matching the cost of single-judge evaluation. Empirical results on MT-Bench and MindEval validate the effectiveness of CyclicJudge as predicted, across both general-purpose and domain-specific evaluation settings.

As AI systems approach superhuman capabilities, scalable oversight increasingly relies on LLM-as-a-judge frameworks where models evaluate and guide each other's training. A core assumption is that binary preference labels provide only semantic supervision about response quality. We challenge this assumption by demonstrating that preference labels can function as a covert communication channel. We show that even when a neutral student model generates semantically unbiased completions, a biased judge can transmit unintended behavioral traits through preference assignments, which even strengthen across iterative alignment rounds. Our findings suggest that robust oversight in superalignment settings requires mechanisms that can detect and mitigate subliminal preference transmission, particularly when judges may pursue unintended objectives.

Large Language Models (LLMs) have revolutionized AI-generated content evaluation, with the LLM-as-a-Judge paradigm becoming increasingly popular. However, current single-LLM evaluation approaches face significant challenges, including inconsistent judgments and inherent biases from pre-training data. To address these limitations, we propose CollabEval, a novel multi-agent evaluation framework that implements a three-phase Collaborative Evaluation process: initial evaluation, multi-round discussion, and final judgment. Unlike existing approaches that rely on competitive debate or single-model evaluation, CollabEval emphasizes collaboration among multiple agents with strategic consensus checking for efficiency. Our extensive experiments demonstrate that CollabEval consistently outperforms single-LLM approaches across multiple dimensions while maintaining robust performance even when individual models struggle. The framework provides comprehensive support for various evaluation criteria while ensuring efficiency through its collaborative design.

Long-form question answering (LFQA) demands nuanced evaluation of multi-sentence explanatory responses, yet existing metrics often fail to reflect human judgment. We present LFQA-HP-1M, a large-scale dataset comprising 1.3M human pairwise preference annotations for LFQA. We propose nine rubrics for answer quality evaluation, and show that simple linear models based on these features perform comparably to state-of-the-art LLM evaluators. We further examine transitivity consistency, positional bias, and verbosity biases in LLM evaluators and demonstrate their vulnerability to adversarial perturbations. Overall, this work provides one of the largest public LFQA preference datasets and a rubric-driven framework for transparent and reliable evaluation.

Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-language rubrics and validated on benchmarks. We identify a previously under-recognized vulnerability in this workflow, which we term Rubric-Induced Preference Drift (RIPD). Even when rubric edits pass benchmark validation, they can still produce systematic and directional shifts in a judge's preferences on target domains. Because rubrics serve as a high-level decision interface, such drift can emerge from seemingly natural, criterion-preserving edits and remain difficult to detect through aggregate benchmark metrics or limited spot-checking. We further show this vulnerability can be exploited through rubric-based preference attacks, in which benchmark-compliant rubric edits steer judgments away from a fixed human or trusted reference on target domains, systematically inducing RIPD and reducing target-domain accuracy up to 9.5% (helpfulness) and 27.9% (harmlessness). When these judgments are used to generate preference labels for downstream post-training, the induced bias propagates through alignment pipelines and becomes internalized in trained policies. This leads to persistent and systematic drift in model behavior. Overall, our findings highlight evaluation rubrics as a sensitive and manipulable control interface, revealing a system-level alignment risk that extends beyond evaluator reliability alone. The code is available at: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface. Warning: Certain sections may contain potentially harmful content that may not be appropriate for all readers.

Existing LLM-as-a-Judge systems suffer from three fundamental limitations: limited adaptivity to task- and domain-specific evaluation criteria, systematic biases driven by non-semantic cues such as position, length, format, and model provenance, and evaluation inconsistency that leads to contradictory judgments across different evaluation modes (e.g., pointwise versus pairwise). To address these issues, we propose FairJudge, an adaptive, debiased, and consistent LLM-as-a-Judge. Unlike prior approaches that treat the judge as a static evaluator, FairJudge models judging behavior itself as a learnable and regularized policy. From a data-centric perspective, we construct a high-information-density judging dataset that explicitly injects supervision signals aligned with evaluation behavior. Building on this dataset, we adopt a curriculum-style SFT-DPO-GRPO training paradigm that progressively aligns rubric adherence, bias mitigation, and cross-mode consistency, while avoiding catastrophic forgetting. Experimental results on multiple internal and public benchmarks show that FairJudge consistently improves agreement and F1, reduces non-semantic biases, and outperforms substantially larger instruction-tuned LLMs. All resources will be publicly released after acceptance to facilitate future research.

Large language models are widely employed as evaluators, a paradigm commonly referred to as LLM-as-a-judge. Prior research has predominantly examined point-wise or pair-wise evaluation protocols; in contrast, our focus is on rubric-based evaluation, which has been attracting increasing attention owing to its utility for training models in domains where verification is otherwise difficult. In this work, we show that rubric-based evaluation implicitly resembles a multiple-choice setting and therefore exhibits position bias: LLMs tend to prefer score options that appear at specific positions within the rubric list. Through controlled experiments across multiple models and datasets, we demonstrate that this position bias is consistent. Its direction, however, is model-specific: some judges favor the first option, while others favor the last. We further identify a second, orthogonal axis of bias: when a prompt scores several criteria simultaneously, the ordering of the criteria itself shifts the resulting scores. We additionally explore permuting the order of the rubric options as a means of mitigating position bias, and find that although the bias can be attenuated, improvements in the correlation between model judgments and human annotations are obtained primarily for models that exhibit strong bias. Our results recast rubric-based LLM-as-a-judge as a multiple-choice problem with measurable, model-specific position bias, and we further confirm that only a small number of random order permutations are sufficient to reduce the error introduced by this bias for the majority of models.

Large language model (LLM) evaluation commonly relies on fixed human-authored benchmarks or centralized human/model judges. Fixed benchmarks can age and enter model training corpora, while centralized judges introduce their own scoring preferences. We cast LLM evaluation as a multi-agent peer-assessment problem and introduce Agents-as-Peers, a distributed evaluation protocol in which every agent generates tasks, answers the shared task set, and judges the answers of its peers, with no privileged evaluator. Self-ratings are excluded from the ranking, and judging is repeated under controlled identity and presentation regimes so that self bias, identity effects, presentation-position sensitivity, and judge severity are measured rather than left implicit. In an instantiation with 12 LLMs and 420 agent-generated questions, Agents-as-Peers aggregates 55,440 non-self peer judgments per regime. Most agents assign their own answers higher scores than those answers receive from peer judges; visible identity changes scores for some agents; the first fixed answer slot is associated with scores 0.39 points higher than the shuffled baseline; and judges differ markedly in severity. On labeled benchmarks, peer scores align strongly with accuracy on TruthfulQA (Pearson r=0.904) and GSM8K (r=0.873). On TruthfulQA, the peer-score correlation is substantially higher than the self-score correlation (r=0.538). A benchmark-overlap screen and item-level human audit further characterize the endogenous task set. These results support Agents-as-Peers as a cohort-relative measurement protocol whose ranking is reported together with the biases and rater effects that shape it. Code and dataset: https://github.com/caura-ai/caura-PeerRank

Recent research has shown that large language models (LLMs) favor their own outputs when acting as judges, undermining the integrity of automated post-training and evaluation workflows. However, it is difficult to disentangle which behaviors are explained by narcissism versus experimental confounds. Specifically, LLM evaluators may deliver self-preferring verdicts when comparing responses to questions they fail on; these verdicts may not depend on the identity of the author, but on evaluator quality. We correct this by directly comparing the judge's voting distribution in cases where it evaluates itself versus another model. This evaluator quality baseline reveals that only 51% of examples in previous findings retain statistical significance against this null hypothesis, covering 89.6% of total self-preference probability mass. Finally, we compare the entropy of voting distributions, suggesting uncertainty-driven overlap, and show that our procedure enables more careful documentation against the backdrop of judge-bias research.

As LLMs rapidly saturate existing benchmarks, automated benchmark creation using LLMs (LLM as a benchmark) where a model generates test inputs (LLM as a testset) and evaluates outputs (LLM as an evaluator) has gained traction as a cheap alternative to human curation. We show that this paradigm has a fundamental problem: LLM-generated benchmarks systematically favor the model that created them. Using machine translation as our primary testbed, we find that self bias arises from two compounding sources, LLM as a testset and LLM as an evaluator, and their combination amplifies the effect. Crucially, even when test data is generated with explicit diversity controls, each modelś implicit stylistic tendencies produce homogeneous, model-specific outputs that inflate its own scores. Increasing source text diversity, using our proposed diversity metric, partially mitigates this bias. Self bias is strong enough to cause each model to rank itself first, overriding the peer consensus ordering. We confirm that the phenomenon extends to open-ended generation on the Chatbot Arena task.

Large language models (LLMs) increasingly serve as automated evaluators, yet they suffer from "self-preference bias": a tendency to favor their own outputs over those of other models. This bias undermines fairness and reliability in evaluation pipelines, particularly for tasks like preference tuning and model routing. We investigate whether lightweight steering vectors can mitigate this problem at inference time without retraining. We introduce a curated dataset that distinguishes self-preference bias into justified examples of self-preference and unjustified examples of self-preference, and we construct steering vectors using two methods: Contrastive Activation Addition (CAA) and an optimization-based approach. Our results show that steering vectors can reduce unjustified self-preference bias by up to 97%, substantially outperforming prompting and direct preference optimization baselines. Yet steering vectors are unstable on legitimate self-preference and unbiased agreement, implying self-preference spans multiple or nonlinear directions. This underscores both their promise and limits as safeguards for LLM-as-judges and motivates more robust interventions.