Language Model Bias Evaluation

We propose a novel perspective for probing LLM sycophancy in a direct and neutral way, mitigating various forms of uncontrolled bias, noise, or manipulative language, deliberately injected to prompts in prior works. A key novelty of our approach is the use of an LLM-as-a-judge in a zero-sum betting game. Within this framework, sycophancy serves one individual (the user) while explicitly incurring cost on another. Comparing 11 leading models we find that while most models exhibit significant sycophantic tendencies in the common setting, in which sycophancy is self-serving to the user and incurs no cost on others, seven of the models exhibit ``moral remorse'', five of which significantly over-compensate for their sycophancy in case it explicitly harms a third party. We refer to this phenomenon as `anti-sycophancy' bias and discuss possible causes for this shift.

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying either on artificial prompts that poorly reflect real-world use or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework that uses named entities as controlled probes to measure systematic disparities in model behavior. Synthetic data enables us to construct diverse, controlled inputs, and we show that it reliably reproduces bias patterns observed in natural text, supporting its use for large-scale analysis. Using this framework, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. We find consistent patterns: models penalize right-wing politicians and favor left-wing politicians, prefer Western and wealthier countries over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not mitigate Western-aligned preferences. These findings highlight the need for systematic bias auditing before deploying LLMs in high-stakes applications. Our framework is extensible to other domains and tasks, and we make it publicly available to support future work.

Large language models are increasingly used to interpret politically contested questions, value-laden material on which there is no single correct answer, only competing interpretive traditions. We ask whether a model's choice among those traditions can turn on the language of the prompt rather than the content. Comparing two frontier models, ChatGPT 5.2 and Claude Opus 4.5, on one contested Ukrainian civil-society document under semantically matched Russian and Ukrainian prompts, we find that both shift along the same axis on identical source text: Russian prompts elicit delegitimizing readings of the document's authors and Ukrainian prompts legitimating ones. The magnitude is model-dependent but neither model is neutral: each adopts a language-dependent stance, and the difference is one of degree. Because contested political questions admit no correct reading against which to measure, we read this as language-conditioned variation in which interpretive tradition a model activates: the model neither holds a single stance nor surfaces the plurality of available ones, but silently adopts the dominant frame of the prompt's language. We draw out the consequences for pluralism-aware evaluation, which must probe the same content across the languages a model serves, and for pluralistic alignment in multilingual settings.

Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach effectively mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.

Large language models increasingly function as artificial reasoners: they evaluate arguments, assign credibility, and express confidence. Yet their responses can leave the epistemic policies governing these evaluations implicit. This paper argues for an epistemic constitution for AI: explicit, contestable meta-norms regulating how systems form and express beliefs. Source attribution provides the motivating case. An exploratory audit suggested that expectations about a source's position intrude on argument evaluation. A preregistered study (arXiv:2609.35286) then found content-dependent effects of source attribution, with selected written evaluations supporting source-position fit as an explanation. The audit also revealed conflicting justifications for attending to sources. Source independence, however, is not a neutral default: in testimonial contexts, a source's position and the costs of speaking against interest can provide relevant evidence. I distinguish two approaches to epistemic constitution design: the Platonic, which mandates formal correctness and default source-independence from a privileged standpoint, and the Liberal, which rejects such privilege and protects conditions for collective inquiry while allowing principled source-attending grounded in epistemic vigilance. I defend the Liberal approach, sketch a constitutional core of eight principles and four orientations, and argue that AI epistemic governance requires explicit, contestable norms for evaluating testimony, responding to evidence, and revising judgements.

People use LLMs for reproductive-health questions, including abortion-related support. A response can sound supportive while answers reinforce harmful assumptions: judgment is likely, secrecy is safer, and support is limited. We introduce behavioral coherence evaluation, a design-time method that uses validation evidence from an established instrument to test relations among outputs. Using the Individual Level Abortion Stigma Scale, we prompted five LLMs to complete questionnaires for 627 personas and reviewed flagged patterns with five reproductive-health experts. Models scored personas lower on self-judgment but higher on worries about judgment; most made worries the highest-scoring dimension, although it was lowest in the ILAS reference sample. Four of five models reversed the reference direction by generating significantly higher worries about judgment scores for Black personas. Models defaulted to extreme secrecy after abortion despite varying stigma patterns across personas. Expert review showed that disclosure guidance requires context about relationship safety, legal risk, and trusted support.

While recent safety guardrails effectively suppress overtly biased outputs, subtler forms of social bias emerge during complex logical reasoning tasks that evade current evaluation benchmarks. To fill this gap, we introduce a new evaluation framework, PRIME (Puzzle Reasoning for Implicit Biases in Model Evaluation), that uses logic grid puzzles to systematically probe the influence of social stereotypes on logical reasoning and decision making in LLMs. Our use of logic puzzles enables automatic generation and verification, as well as variability in complexity and biased settings. PRIME includes stereotypical, anti-stereotypical, and neutral puzzle variants generated from a shared puzzle structure, allowing for controlled and fine-grained comparisons. We evaluate multiple model families across puzzle sizes and test the effectiveness of prompt-based mitigation strategies. Focusing our experiments on gender stereotypes, our findings highlight that models consistently reason more accurately when solutions align with stereotypical associations. This demonstrates the significance of PRIME for diagnosing and quantifying social biases perpetuated in the deductive reasoning of LLMs, where fairness is critical.

As large language models (LLMs) are adopted into frameworks that grant them the capacity to make real decisions, it is increasingly important to ensure that they are unbiased. In this paper, we argue that the predominant approach of simply removing existing biases from models is not enough. Using a paradigm from the psychology literature, we demonstrate that LLMs can spontaneously develop novel social biases about artificial demographic groups even when no inherent differences exist. These biases result in highly stratified task allocations, which are less fair than assignments by human participants and are exacerbated in newer and larger models. In humans, emergent biases like these have been shown to result from exploration-exploitation trade-offs, where the decision-maker explores too little, allowing early observations to strongly influence impressions about entire demographic groups. To alleviate this effect, we explore a series of interventions targeting model inputs, problem structure, and explicit steering. While most interventions have limited effect, explicitly incentivizing exploration robustly reduces stratification, highlighting the need for better multifaceted objectives to mitigate bias. These results reveal that LLMs are not merely passive mirrors of human social biases, but can actively create new ones from experience, raising urgent questions about how these systems will shape societies over time.

Large language models (LLMs) are now widely deployed in user-facing applications, reaching hundreds of millions of users worldwide. Despite their widespread adoption, growing reliance on their outputs raises significant concerns, particularly as users may be exposed to model-inherent biases that disadvantage or stereotype certain groups. However, existing bias benchmarks commonly rely on simple templated prompts or restrictive multiple-choice questions that fail to capture the complexity of real-world user interactions. In this work, we address this gap by introducing a counterfactual framework that automatically generates realistic, open-ended questions for LLM bias evaluation. Through iterative question mutation, our approach systematically explores areas where models are most likely to exhibit biased behavior. Beyond just detecting harmful biases, we also capture increasingly relevant response dimensions, such as asymmetric refusals and explicit bias acknowledgment. Building on this, we construct CAB, a diverse and human-verified benchmark for realistic and nuanced bias evaluations on current frontier LLMs. Our evaluation using CAB highlights the continued need for fairness research by showing that all examined models exhibit persistent biases across certain scenarios.

As Large Language Models (LLMs) develop stronger multilingual capabilities, their sensitivity to culturally diverse entities becomes increasingly important. Prior work by Naous et al. (2024) has shown that LLMs often favor Western-associated entities in Arabic. Due to the lack of entity-centric multilingual benchmarks, it remains unclear if such biases also manifest in various non-Western languages. In this paper, we introduce Camellia, a benchmark for evaluating entity-centric cultural biases in nine Asian languages, spanning six Asian cultures. Camellia includes 19,530 manually annotated entities associated with the covered Asian or Western cultures, as well as 2,173 masked contexts for these entities derived from social media posts. Using Camellia, we evaluate cultural biases in four recent multilingual LLMs across three tasks: cultural context adaptation, sentiment association, and entity extractive QA. Our analyses show that LLMs struggle with cultural adaptation across these languages, with performance differing across models developed in different regions. We further observe that different LLM families can hold distinct biases, reflected in the ways they link cultures to particular sentiments. Lastly, we find that LLMs can struggle with context understanding in some Asian languages, creating performance gaps between cultures in entity extraction.

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.

Large language models (LLMs) are increasingly deployed in politically sensitive contexts, raising concerns about their susceptibility to ideological biases. In this work, we examine how synthetic persona conditioning shapes ideological expression across seven open-weight instruction-tuned models (7B-72B parameters) using the Political Compass Test (62 statements) as a standardized behavioral probe. Across three studies involving 200,000 synthetic personas and more than 260 million model responses, we analyze implicit and explicit malleability, as well as theme-associated variations. We find that: (i) larger models exhibit broader implicit ideological coverage, increasing from 14-35% for 7-8B models to up to 49% for 70B+ models; (ii) explicit ideological priming induces large and statistically significant shifts, with right-authoritarian cues moving all models in the intended direction and producing larger effects in most model-axis comparisons; (iii) left-libertarian priming produces more heterogeneous responses, including counter-directional economic shifts in three of four 7-8B models, while all 70B+ models move in the intended direction; and (iv) theme-associated semantic content in persona descriptions is linked to systematic and interpretable directional shifts in ideological space. While our results identify an upstream mechanism through which persona conditioning can alter model responses under a standardized ideological probe, we do not test whether such shifts affect users beliefs, decisions, or political behavior. Our findings are best understood as evidence of ideological malleability at the generation layer, highlighting the need to account for interactional factors when evaluating political neutrality, fairness, and safety in English-prompted, persona-conditioned language models.

Understanding biases and stereotypes encoded in the weights of Large Language Models (LLMs) is crucial for developing effective mitigation strategies. However, biased behavior is often subtle and non-trivial to isolate, even when deliberately elicited, making systematic analysis and debiasing particularly challenging. To address this, we introduce a simple, cost-effective, and generalizable framework \texttt{BiasGym} for reliably injecting, analyzing, and mitigating conceptual associations of biases within LLMs. \texttt{BiasGym} consists of two modules: \texttt{Inject}, which injects specific biases into the model via token-based fine-tuning while keeping the model frozen, followed by two debiasing methods that leverage these injected signals to identify and reliably suppress (\texttt{Scope}) or \texttt{Steer} the components responsible for biased behavior. Our framework enables consistent bias elicitation for better localization of bias conceptual association in the model space, supports targeted debiasing without degrading performance on downstream tasks, and generalizes to biases unseen during fine-tuning. We demonstrate the effectiveness of our proposed framework in reducing real-world stereotypes (e.g., people from Italy being `reckless drivers'), showing its utility for both safety interventions and interpretability research.

Instruct-based large language models (LLMs) have been shown to propagate and even amplify gender bias when prompted with contextually constrained instructions (e.g., writing a text from a description or selecting a gendered pronoun). However, little attention has been paid to biases in responses to contextually unconstrained (generic) instructions conveyed by gendered language, particularly masculine generics (MG). MG, found in many gender-marked languages, denote the use of the masculine gender as a supposedly neutral reference to mixed-gender groups or individuals whose gender is unknown or non-binary. Yet, psycholinguistic studies demonstrate that MG are not neutral and systematically induce gender bias. This study investigates how both local and proprietary LLMs are MG-biased when responding to generic prompts in French, examining LLMs' MG bias rates and use of gender-fair language (GFL). We create a 16k+ human noun database from existing lexical resources and evaluate six LLMs on four instruction-response datasets under two conditions: prompts with and without MG. Overall, we find that $\approx$27.57% of LLMs' responses to MG-filtered generic instructions are MG-biased ($\approx$78.55% with MG-containing prompts). Moreover, we find that LLMs rarely use GFL spontaneously. These findings highlight the persistence of MG bias in LLM outputs and models' limited tendency towards GFL strategies.

Large language models (LLMs) are increasingly used in high-stakes decisions such as hiring and college admissions, making their social bias a critical concern. While LLMs are trained to refuse explicitly biased requests, bias can be leaked implicitly during LLM planning and reasoning process. As code becomes the primary medium for LLM internal logic-writing, we introduce FairCoder, a benchmark that frames decision-making as coding tasks to systematically probe LLM bias across employment, education, and healthcare domains, covering multiple fairness definitions. Considering that existing metrics may fail when LLMs frequently refuse the request, we propose FairScore, a metric that jointly captures refusal behavior and group-level outcome diversity. Experiments with a 1k-sample dataset on powerful LLMs reveal consistent and previously underexplored bias patterns, such as prioritizing applicants from high-income families in college admissions. Our findings highlight the risks of deploying LLMs as decision-making agents and provide a comprehensive evaluation framework for future research.

Large language models (LLMs) reproduce homogeneity bias -- the tendency to portray marginalized groups as more internally similar than dominant groups -- but whether this bias is stable or an artifact of inference settings has only been studied in single proprietary models. We map homogeneity bias across a 5x5 temperature-by-top-p grid in seven open-weight instruction-tuned LLMs (7-20B parameters). Hispanic and Asian Americans are portrayed as more homogeneous than White Americans in at least 18 of 20 hyperparameter configurations across six of seven models, including at extreme sampling settings. African American and gender bias show model-specific variation in direction. A conservative cell-level re-analysis confirms Hispanic and Asian homogeneity as robust, while weaker African American and gender signals largely do not survive, establishing group-specific robustness. We also apply the same grid to a names-based paradigm in which group identity is signaled via racially distinctive surnames rather than explicit labels. The names paradigm corroborates Hispanic and Asian homogeneity bias, but Black-coded surnames elicit robustly less homogeneous outputs than White-coded names in every model tested -- a reversal absent from the label paradigm -- showing that how group identity is operationalized shapes which biases surface and in which direction.

Large language models (LLMs) may encode biased associations from heterogeneous training corpora that are not immediately visible under ordinary prompting, but can surface when the model is steered toward particular demographic personas. Such behavior often manifests not as explicit toxic output, but as systematic performance differences across semantically equivalent tasks, making the resulting bias difficult to detect and mitigate. To address this issue, we formalize the implicit bias problem as persona-induced performance disparity and argue that bias evidence should be treated as a graded signal rather than a binary label. Motivated by this observation, we model biased knowledge as a fuzzy subset equipped with an explicit membership function that reflects the strength of bias evidence for each candidate example. Building on this formulation, we propose Bayesian-Theory-based Bias Removal (BTBR), a hybrid probabilistic-fuzzy framework for identifying and removing latent bias traces from model parameters. BTBR first performs likelihood-ratio screening to measure how strongly candidate samples align with a target biased persona, then converts high-membership samples into structured knowledge triples, and finally applies targeted model editing with a lightweight fuzzy rule scheduler to reduce collateral performance degradation under high entanglement risk. Extensive experiments across multiple bias sources, tasks, model families and editing backends show that BTBR consistently reduces persona-induced performance gaps while preserving general reasoning ability. These results demonstrate that combining probabilistic evidence with fuzzy degree modeling provides an effective and practical approach for mitigating implicit bias in large language models.

Warning: This paper contains content that may be offensive or upsetting. There has been a significant increase in the usage of large language models (LLMs) in various applications, both in their original form and through fine-tuned adaptations. As a result, LLMs have gained popularity and are being widely adopted by a large user community. However, one of the concerns with LLMs is the potential generation of socially biased content. The existing evaluation methods have many constraints, and their results exhibit a limited degree of interpretability. In this work, we propose a bias evaluation framework named GPTBIAS that leverages the high performance of LLMs (e.g., GPT-4 \cite{openai2023gpt4}) to assess bias in models. We also introduce prompts called Bias Attack Instructions, which are specifically designed for evaluating model bias. To enhance the credibility and interpretability of bias evaluation, our framework not only provides a bias score but also offers detailed information, including bias types, affected demographics, keywords, reasons behind the biases, and suggestions for improvement. We conduct extensive experiments to demonstrate the effectiveness and usability of our bias evaluation framework.

During training, Large Language Models (LLMs) learn social regularities that can lead to gender bias in downstream applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated on structured benchmarks, which raises two concerns: output-level evaluation does not reveal whether alignment modifies the model's underlying representations, and structured benchmarks may not reflect realistic usage scenarios. We propose a unified framework to jointly analyze intrinsic and extrinsic gender bias in LLMs using identical neutral prompts, enabling direct comparison between gender-related information encoded in internal representations and bias expressed in generated outputs. Contrary to prior work reporting weak or inconsistent correlations, we find a consistent association between latent gender information and expressed bias when measured under the unified protocol. We further examine the effect of alignment through supervised fine-tuning aimed at reducing gender bias. Our results suggest that while the latter indeed reduces expressed bias, measurable gender-related associations are still present in internal representations, and can be reactivated under adversarial prompting. Finally, we consider two realistic settings and show that debiasing effects observed on structured benchmarks do not necessarily generalize, e.g., to the case of story generation.

We systematically compare word order preferences in decoder-only language models across 192 artificial languages and typologically diverse natural languages. On artificial languages, models exhibit a left-branching preference that aligns with neither natural language universals nor human word order learning biases. On natural languages, monolingual models show no clear base word order bias at small scales, but as data grows, a preference for right-branching subject-verb-object (SVO) languages emerges while SOV falls behind despite being the most frequent order cross-linguistically. This SVO advantage extends to multilingual models and correlates with language resource level and data quality rather than word order. Thus, the same architecture exhibits opposite preferences on artificial and natural languages, establishing that word order biases observed in practice are data-driven. Since highly-resourced languages are overwhelmingly SVO, these biases risk gradually reducing word order diversity, particularly in languages that productively use multiple word orders, with the widespread adoption of LLMs.