Language Model Bias Evaluation

Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the "Explain this post" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.

Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes, leaving judge explanations underexplored. We instead ask whether LLM judges are cue-invariant, i.e., whether their rankings and explanations remain stable when non-evidential cues are perturbed while holding the underlying texts fixed. We introduce a suite of cue interventions (Blind, Truth, Flip, Placebo, Reveal-After) and tie-aware metrics that quantify outcome anchoring and rationale anchoring, including label-aligned rhetoric and explanation drift, alongside consistency and stereotype-intrusion checks. We design anchoring attacks using verbosity and confidence cues, and compare two mitigations: structured chain-of-thought prompting and PROOF-BEFORE-PREFERENCE (evidence lock, score, rank). Using a new dataset of 1,000 summaries from traditional extractive models and LLMs, we find substantial cue-anchored rationalization under label and placebo perturbations, while PROOF-BEFORE-PREFERENCE markedly improves cue invariance over baselines.

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

Large language models (LLMs) exhibit systematic preferences for well-known entities, a phenomenon often attributed to popularity bias. However, the extent to which these preferences reflect real-world popularity versus statistical exposure during pretraining remains unclear, largely due to the inaccessibility of most training corpora. We provide the first direct, large-scale analysis of popularity bias grounded in fully observable pretraining data. Leveraging the open OLMo models and their complete pretraining corpus, Dolma, we compute precise entity-level exposure statistics across 7.4 trillion tokens. We analyze 2,000 entities spanning five types (Person, Location, Organization, Art, Product) and compare pretraining exposure against Wikipedia pageviews and two elicited LLM popularity signals: direct scalar estimation and pairwise comparison. Our results show that pretraining exposure strongly correlates with Wikipedia popularity, validating exposure as a meaningful proxy for real-world salience during the training period. More importantly, we find that LLM popularity judgments align more closely with exposure than with Wikipedia, especially when elicited via pairwise comparisons. This alignment is strongest for larger models and persists in the long tail, where Wikipedia popularity becomes unreliable. Overall, our findings demonstrate that popularity priors in LLMs are primarily shaped by pretraining statistics rather than external popularity signals, offering concrete evidence that data exposure plays a central role in driving popularity bias.

Recent works have analyzed the impact of individual components of neural networks on gendered predictions, often with a focus on mitigating gender bias. However, mechanistic interpretations of gender tend to (i) focus on a very specific gender-related task, such as gendered pronoun prediction, or (ii) fail to distinguish between the production of factually gendered outputs (the correct assumption of gender given a word that carries gender as a semantic property) and gender biased outputs (based on a stereotype). To address these issues, we curate \gknow, a benchmark to assess gender knowledge and gender bias in language models across different types of gender-related predictions. \gknow allows us to identify and analyze circuits and individual neurons responsible for gendered predictions. We test the impact of neuron ablation on benchmarks for disentangling stereotypical and factual gender (DiFair and the test set of GKnow), as well as StereoSet. Results show that gender bias and factual gender are severely entangled on the level of both circuits and neurons, entailing that ablation is an unreliable debiasing method. Furthermore, we show that benchmarks for evaluating gender bias can hide the decrease in factual gender knowledge that accompanies neuron ablation. We curate GKnow as a contribution to the continuous development of robust gender bias benchmarks.

Large language model (LLM)-based simulations of clinical patients are increasingly used for research and training, yet their validity requires persona stability: coherent maintenance of an assigned psychological profile across and within conversations. We evaluate this prerequisite using eating disorder personas grounded in five published case vignettes, a dual-assessment framework (self-report + independent observer ratings), and validated psychometric instruments (EDE-Q) with known ground-truth scores. Across six LLMs and two experiments (between-conversation stability (Exp. I) and within-conversation stability (Exp. II)), we find that LLMs are paradoxically too stable and too inaccurate: variability is negligible, yet all models systematically overshoot ground-truth severity by 12-30% of the scale range (0.7-1.8 points on a 0-6 scale). The mechanism is selective stereotyping: models differentiate cases on behavioural items (dietary restraint) but maximise cognitive-affective items (body dissatisfaction, weight preoccupation) at ceiling regardless of case severity. Additional conversational context does not improve accuracy; it compounds the overshoot. LLMs can portray severe eating pathology but lack a representation of moderate clinical presentations, a "missing middle".

LLM annotators are increasingly used in computational social science (CSS), but it is unclear whether their alignment-shaped errors preserve the empirical conclusions a researcher would report. We audit three open-source 7B instruction-tuned models (Zephyr, Mistral-Instruct, Qwen2.5-Instruct) across six TweetEval tasks under four prompt conditions (72 cells) and find that social-desirability failures do not run in a single direction. Zephyr exhibits leniency bias, systematically under-applying harmful labels (offensive language: false benign rate 0.729, false alarm rate 0.031). Mistral and Qwen exhibit overcorrection, over-applying the same labels (Mistral hate-speech FAR = 0.604). All three models exhibit neutrality bias on abortion stance, underestimating opposition prevalence by 24 to 40 percentage points and inflating the neutral label. None of the four prompting interventions we test (neutral, safety framing, depersonalized, chain-of-thought) corrects these failures across models; safety framing can worsen stance distortion. Strikingly, Zephyr's hate-speech prevalence estimate matches the gold rate exactly while its class-conditional errors are large in both directions, an accidental cancellation that misleads aggregate validation. We translate these patterns into a three-part taxonomy with diagnostic FBR/FAR signatures and a lightweight gold-sample validation protocol. The headline for trustworthy CSS: a model that looks calibrated on aggregate metrics can still flip the substantive empirical conclusion a researcher would report.

Automated systems built on artificial intelligence (AI) are increasingly deployed across high-stakes domains, raising critical concerns about fairness and the perpetuation of demographic disparities that exist in the world. In this context, causal inference provides a principled framework for reasoning about fairness, as it links observed disparities to underlying mechanisms and aligns naturally with human intuition and legal notions of discrimination. Prior work on causal fairness primarily focuses on the standard machine learning setting, where a decision-maker constructs a single predictive mechanism $f_{\widehat Y}$ for an outcome variable $Y$, while inheriting the causal mechanisms of all other covariates from the real world. The generative AI setting, however, is markedly more complex: generative models can sample from arbitrary conditionals over any set of variables, implicitly constructing their own beliefs about all causal mechanisms rather than learning a single predictive function. This fundamental difference requires new developments in causal fairness methodology. We formalize the problem of causal fairness in generative AI and unify it with the standard ML setting under a common theoretical framework. We then derive new causal decomposition results that enable granular quantification of fairness impacts along both (a) different causal pathways and (b) the replacement of real-world mechanisms by the generative model's mechanisms. We establish identification conditions and introduce efficient estimators for causal quantities of interest, and demonstrate the value of our methodology by analyzing race and gender bias in large language models across different datasets.

Large language models (LLMs) trained on web-scale corpora can memorize sensitive training data, posing significant privacy risks. Differential privacy (DP) has emerged as a principled framework that limits the influence of individual data points during training, yet the relationship between differential privacy and social bias in LLMs remains poorly understood. To investigate this, we present a systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing a DP model against non-DP baselines across four complementary paradigms: sentence scoring, text completion, tabular classification, and question answering. We find that DP reduces bias in sentence scoring tasks, where bias is measured through controlled likelihood comparisons, yet this improvement does not generalize across all tasks. Our results reveal a discrepancy between logit-level bias and output-level bias. Moreover, decreasing memorization does not necessarily reduce unfairness, underscoring the importance of multi-paradigm evaluation when assessing fairness in LLMs.

Multimodal large language models (LLMs) are increasingly explored as automated evaluators in clinical settings, yet their scoring behavior on ordinal clinical scales remains poorly understood. We benchmark three frontier LLM families against supervised deep learning models for scoring Clock Drawing Test (CDT) images on two public datasets using the Shulman rubric. While fully fine-tuned Vision Transformers achieve the best calibration (MAE 0.52, within-1 accuracy 91%), zero-shot LLMs remain competitive on tolerance-based agreement (GPT-5 MAE 0.67, within-1 accuracy 92%) despite higher absolute error. However, per-score analysis reveals that all three LLM families exhibit a pronounced central tendency effect (systematic endpoint compression): predictions are systematically compressed toward the middle of the scale, with over-prediction at the low end (score 0 to 1) and under-prediction at the high end (score 5 to 4). This effect disproportionately affects the clinically critical extremes where accurate scoring most impacts screening decisions for cognitive impairment. Targeted ablations show that neither few-shot exemplars spanning the full score range nor removing clinical terminology from the prompt eliminates the effect. Our findings extend the LLM-as-a-judge bias literature from NLP evaluation to clinical assessment, and highlight the need for calibration-aware evaluation and post-hoc calibration before deploying LLM-based raters in high-stakes screening workflows.

The rapid adoption of LLMs in both research and industry highlights the challenges of deploying them safely and reveals a gap in the systematic evaluation of toxicity benchmarks. As organizations increasingly rely on these benchmarks to certify models for customer-facing applications and automated moderation, unrecognized evaluation biases could lead to the deployment of vulnerable or unsafe systems. This work investigates the robustness of established benchmarking setups and examines how to measure currently neglected intrinsic biases, such as those related to model choice, metrics, and task types. Our experiments uncover significant discrepancies in benchmark behaviors when evaluation setups are altered. Specifically, shifting the task from text completion to summarization increases the tendency of benchmarks to flag content as harmful. Additionally, certain benchmarks fail to maintain consistent behavior when the input data domain is changed. Furthermore, we observe model-specific instabilities, demonstrating a clear need for more robust and comprehensive safety evaluation frameworks.

Multilingual studies of social bias in open-ended LLM generation remain limited: most existing benchmarks are English-centric, template-based, or restricted to recognizing pre-specified stereotypes. We introduce StereoTales, a multilingual dataset and evaluation pipeline for systematically studying the emergence of social bias in open-ended LLM generation. The dataset covers 10 languages and 79 socio-demographic attributes, and comprises over 650k stories generated by 23 recent LLMs, each annotated with the socio-demographic profile of the protagonist across 19 dimensions. From these, we apply statistical tests to identify more than 1{,}500 over-represented associations, which we then rate for harmfulness through both a panel of humans (N = 247) and the same LLMs. We report three main findings. \textbf{(i)} Every model we evaluate emits consequential harmful stereotypes in open-ended generation, regardless of size or capabilities, and these associations are largely shared across providers rather than isolated misbehaviors. \textbf{(ii)} Prompt language strongly shapes which stereotypes appear: rather than transferring as a shared set of biases, harmful associations adapt culturally to the prompt language and amplify bias against locally salient protected groups. \textbf{(iii)} Human and LLM harmfulness judgments are broadly aligned (Spearman $ρ=0.62$), with disagreements concentrating on specific attribute classes rather than specific providers. To support further analyses, we release the evaluation code and the dataset, including model generations, attribute annotations, and harmfulness ratings.

Bias audits of large language models now operate within governance frameworks such as the EU AI Act, making benchmark reliability a security concern in its own right. Many current benchmarks, however, collapse bias into a single scalar from one prompt format and one surface label. This design misses two failure modes that can be exploited without changing model weights. Across prompts, meaning-preserving format changes shift bias endorsement by more than $0.7$ on a fixed statement pool. Within a response, the discrete Selection and free-text Elaboration can take opposing stances, so an apparently clean aggregate may hide substantial internal inconsistency (a ``cancellation trap''). Selection-only and elaboration-only rankings are therefore nearly uncorrelated across eight LLMs (Spearman $ρ= 0.238$, $p = 0.570$): LLaMA3-70B ranks in the middle under selection-only scoring but highest under elaboration-only scoring on the same responses. We introduce \textsc{BiAxisAudit}, a protocol that reports each bias score together with a reliability estimate on two orthogonal axes. The across-prompt axis evaluates each statement under a factorial grid of task format, perspective, role, and sentiment, treating bias as a distribution rather than a point estimate. The within-response axis uses Split Coding to recover Selection and Elaboration as separate signals, measured by the Inconsistency Rate and Divergence Net Imbalance. Across eight LLMs with $80{,}200$ coded responses each, task format alone explains as much variance as model choice; $63.6\%$ of pooled bias signals (up to $85.2\%$ per model) appear in only one coding layer, and prompt-dimension interactions exceed main effects. The instrument also separates real bias reductions from apparent reductions caused by cross-layer redistribution: some prompt configurations reduce both BER and IR, whereas others suppress only selection-layer bias.

Large language models (LLMs) are increasingly deployed not only to make decisions but to explain them. While AI decision fairness has been studied extensively, the fairness of AI explanations (whether LLMs justify decisions with equal quality, depth, tone, and linguistic sophistication across demographic groups) has received little attention. This paper introduces the Explanation Fairness Taxonomy (EFT), a framework comprising five formally defined, operationalizable dimensions: Verbosity Disparity, Sentiment Disparity, Epistemic Hedging Disparity, Decision-Linked Explanation Disparity, and Lexical Complexity Disparity. The taxonomy is instantiated in a controlled empirical study across 80 prompt templates, four consequential decision domains (hiring, medical triage, credit assessment, legal judgment), and five LLMs: GPT-4.1, Claude Sonnet, LLaMA 3.3 70B, GPT-OSS 120B, and Qwen3 32B. Two novel black-box metrics are introduced: the Hedging Density Score (HDS) and the Explanation Faithfulness Proxy (EFP), a heuristic indicator of decision-linked explanation variation. Across up to 400 prompt pairs, all eight EFT metrics show statistically significant disparities (Cohen's d ranging from small to large, all p_BH < 10^(-62)). Model choice is strongly associated with disparity magnitude: Qwen3 32B exhibits verbosity disparities 5.9x larger than LLaMA 3.3 70B. Two prompting-based mitigations show significant reductions in EFP disparity (78-95%) but no significant effect on stylistic dimensions, consistent with the hypothesis that stylistic explanation inequalities are encoded in pre-training distributions and are not resolvable through deployment-level instruction alone. A reproducible measurement framework is offered for explanation-level fairness auditing, with implications for AI regulation and deployment practice.

Since the advent of Large Language Models (LLMs), a significant area of research has focused on their intrinsic biases, particularly in political discourse. This study investigates a different but related concept, "political plasticity", which is defined as the capacity of models to adapt their responses based on the user supplied context. To analyze this, a testing framework was developed using an expanded corpus of 200 politically-oriented questions across economic and personal freedom axes, based on a prior framework by Lester (1996). The study explored several methods to induce political bias, including simplified and topic-based system prompts, as well as user prompts with few-shot examples. The results show that while system prompts were largely ineffective, user prompts successfully elicited significant ideological shifts, particularly along the Economic Freedom axis in larger and newer models. Through a validation experiment, we examined whether models answer questionnaires by recognizing the underlying question format. Inverting the sense of the questions revealed unexpected, counter-intuitive shifts in most models, suggesting potential data leakage. Finally, we also analyzed how model plasticity varies when the experiment is conducted in different languages. The results reveal subtle yet notable shifts across each of the analyzed languages. Overall, our results indicate that small and older LLMs exhibit limited or unstable political plasticity, whereas newer frontier models display reliable, expected adaptability.

Large language models (LLMs) are increasingly deployed as autonomous agents in offensive cybersecurity. In this paper, we reveal an interesting phenomenon: different agents exhibit distinct attack patterns. Specifically, each agent exhibits an attack-selection bias, disproportionately concentrating its efforts on a narrow subset of attack families regardless of prompt variations. To systematically quantify this behavior, we introduce CyBiasBench, a comprehensive 630-session benchmark that evaluates five agents on three targets and four prompt conditions with ten attack families. We identify explicit bias across agents, with different dominant attack families and varying entropy levels in their attack-family allocation distributions. Such bias is better characterized as a trait of the agents, rather than a factor associated with the attack success rate. Furthermore, our experiments reveal a bias momentum effect, where agents resist explicit steering toward attack families that conflict with their bias. This forced distribution shift does not yield measurable improvements in attack performance. To ensure reproducibility and facilitate future research, we release an interactive result dashboard at https://trustworthyai.co.kr/CyBiasBench/ and a reproducibility artifact with aggregated session-level statistics and full evaluation scripts at https://github.com/Harry24k/CyBiasBench.

Gender bias in large language models has primarily been investigated for English, while languages with grammatical or morphological gender remain comparatively understudied. This paper investigates how and when gender information emerges in a Dutch BERT model trained from scratch, offering one of the first checkpoint-level analyses of bias formation in a Transformer architecture for a language combining overt morphological gender marking and generic forms. By extracting contextual embeddings throughout training, we construct dynamic gender subspaces using linear SVMs to trace when gender becomes linearly encoded and how this encoding evolves over time. Contextual embeddings are often assumed to integrate contextual cues robustly, allowing models to adjust the representation of a word depending on its more local usage. We therefore test whether explicit gender cues in controlled sentence templates (e.g., Zij is een loodgieter ('She is a plumber')) can override learned statistical associations (plumber -> male). Our findings challenge this assumption: although gender becomes clearly linearly separable around epoch 20 and is distributed across multiple embedding dimensions, the model struggles to update its internal gender representation in light of explicit contextual cues in short sentence templates. Stereotypical gender-profession pairings are predicted far more accurately than anti-stereotypical ones, and generic forms in Dutch systematically default to a male interpretation, even when the context explicitly denotes a female referent. Together, our results seem to indicate that contextualization in the representations learned by our Dutch BERT model is not sufficiently dynamic along the probed gender direction: explicit gender cues in anti-stereotypical contexts are not reliably reflected in the resulting representations, resulting in persistent male-default behaviour.

Qualitative analysis plays a pivotal role in understanding the human and social aspects of software engineering. However, it remains a demanding process shaped by the subjective interpretation of individual researchers and sensitive to methodological choices such as prompt design. Recent advancements in Large Language Models (LLMs) offer promising opportunities to support this type of analysis, although their reliability in reproducing human qualitative reasoning under varying prompting conditions remains largely untested. This study presents a controlled empirical evaluation of three LLMs -- Claude Haiku, DeepSeek-Chat, and Gemini 2.5 Flash -- across two prompt engineering strategies (zero-shot and multi-shot closed coding), using Cohen's kappa as the primary agreement metric over ten independent runs per configuration. Results suggest that multi-shot prompting significantly improves agreement for Claude Haiku (Delta kappa = +0.034, Wilcoxon p = 0.004) but not for DeepSeek-Chat or Gemini 2.5 Flash. Intra-model stability varies substantially -- DeepSeek-Chat and Claude Haiku exhibit the lowest variance (SD approx. 0.017), while Gemini 2.5 Flash is the least stable (SD = 0.038). A systematic over-prediction of "Sharing Negative Feedback" is identified across all models (bias ratios up to 5.25x), alongside consistent under-prediction of "Expressing Concerns." Collectively, these findings provide empirical evidence for prompt engineering guidelines in LLM-assisted qualitative coding for software engineering research.

Annotation automation via Large Language Models (LLMs) is the core approach for scaling NLP datasets; however, LLM behavior with respect to closed-set instructions in low-resource languages has not been well studied. We present MultiSoc-4D, a Bengali social media dataset benchmark, which contains 58K+ social media comments from six sources annotated along four dimensions: category, sentiment, hate speech, and sarcasm. By employing a structured pipeline where ChatGPT, Gemini, Claude, and Grok individually annotate separate partitions, while sharing a common validation set of 20%, we diagnose LLM behavior systematically. We discover a prevalent phenomenon called "instruction-induced label collapse", wherein LLMs show a systematic preference towards fallback labels (Other, Neutral, No), leading to high agreement rates but under-detection of minority categories. For example, we find that LLMs failed to detect 79% and 75% of instances with hateful and sarcastic content compared to a human-calibrated reference. Furthermore, we prove that it represents a "label agreement illusion", statistically validated via almost null Fleiss' Kappa ($κ\approx -0.001$) on sarcasm detection. Across 40+ LLMs, we benchmark this annotation bias propagation within the training pipeline, regardless of architectural differences. We release MultiSoc-4D as a diagnostic benchmark for annotation biases in Bengali NLP.

As LLMs enter conflict monitoring, understanding systematic distortions in their outputs is critical for humanitarian accountability. We evaluate four vanilla open-weight models Gemma 3 4B, Llama 3.2 3B, Mistral 7B, and OLMo 2 7B and two domain-adapted models, AfroConfliBERT and AfroConfliLLAMA, on Nigeria and Cameroon conflict-event classification against ACLED, a gold-standard dataset with multi-stage verification. We find a bifurcated divergence in normative directionality. Open-weight models exhibit statistically significant False Illegitimation bias: Gemma misclassifies to 18.29% of legitimate battles as civilian-targeted violence while making zero False Legitimation errors. By contrast, AfroConfliBERT and AfroConfliLLAMA achieve near-directional neutrality, with Legitimization Bias differences indistinguishable from zero. Yet domain adaptation does not eliminate actor-based selection bias. Both adapted models show statistically significant actor bias comparable to vanilla LLMs; in Nigeria, state actors are legitimized 36.5% more often than non-state actors in identical tactical contexts. Open-weight outputs are also fragile to geography-specific lexical framing: delegitimizing phrases produce flip rates up to 66.7% in Cameroon and 34.2% in Nigeria, while perturbations salient in one context may not matter in another. Error trace profiling shows models mask normative bias through unfaithful rationale confabulations. In contrast, AfroConfliBERT and AfroConfliLLAMA are largely robust, with near-zero flip rates across perturbation categories. Overall, current models are not ready for unsupervised deployment in conflict monitoring. We call for fairness-aware fine-tuning to reduce actor-based selection bias, mandatory adversarial robustness evaluation against lexical manipulation, and context-specific human-in-the-loop oversight calibrated to regional difficulty.

Emergency department triage assigns patients an acuity score that determines treatment priority, and clinical evidence documents persistent gender disparities in human acuity assessment. As hospitals pilot large language models (LLMs) as triage decision support, a critical question is whether these models reproduce or mitigate known biases. We present EQUITRIAGE, a fairness audit of LLM-based ESI assignment evaluating five models (Gemini-3-Flash, Nemotron-3-Super, DeepSeek-V3.1, Mistral-Small-3.2, GPT-4.1-Nano) across 374,275 evaluations on 18,714 MIMIC-IV-ED vignettes under four prompt strategies. Of 9,368 originals, 9,346 are paired with a gender-swapped counterfactual. All five models produced flip rates above a pre-registered 5% threshold (9.9% to 43.8%). Two showed directional female undertriage (DeepSeek F/M 2.15:1, Gemini 1.34:1); two were near-parity; one had high sensitivity with weak male-direction asymmetry. DeepSeek's directional bias coexisted with a low outcome-linked calibration gap (0.013 against MIMIC-IV admission), a Chouldechova-style dissociation between within-group calibration and between-pair counterfactual invariance. Demographic blinding reduced Gemini's flip rate to 0.5%; an age-preserving blind variant left DeepSeek with residual F/M 1.25, implicating age as a residual channel. Chain-of-thought prompting degraded accuracy for all five models. A two-model ablation reveals opposite underlying mechanisms for the same directional phenotype: in Gemini the signal is emergent in the combined name+gender swap, while in DeepSeek the gender token alone carries it. EQUITRIAGE shows that group parity, counterfactual invariance, and gender calibration are distinct fairness properties, that intervention effectiveness is model-dependent, and that per-model counterfactual auditing should precede clinical deployment.

Large language models (LLMs) are increasingly integrated into legal workflows. However, existing benchmarks primarily address proxy tasks, such as bar examination performance or classification, which fail to capture the performance and risks inherent in day-to-day judicial processes. To address this, we publicly release TriBench-Ko, a Korean benchmark designed to evaluate potential deployment risks of LLMs within the context of verified judicial task requirements. It covers four core tasks: jurisprudence summarization, precedent retrieval, legal issue extraction, and evidence analysis. It jointly assesses model behavior across multiple deployment risk categories, including inaccuracy (hallucination, omission, statutory misapplication), biases (demographic, overcompliance), inconsistencies (prompt sensitivity, non-determinism), and adjudicative overreach. Each item is structured to systematically assess both task performance and a specific risk type based on real judicial decisions. Our evaluation of a range of contemporary LLMs reveals that many models frequently manifest significant risks, most notably struggling with precedent retrieval and failing to capture critical legal information. We provide a comprehensive diagnosis of these LLMs and pinpoint critical areas where LLM-generated outputs in judicial contexts necessitate rigorous inspection and caution. Our dataset and code are available at https://github.com/holi-lab/TriBench-Ko

Large language models (LLMs) are widely recognised for their applications in natural language generation and are increasingly used for code generation tasks. However, concerns about bias in their generated outputs remain significant. This paper focuses on GPT-4o and Gemini, mainstream tools for code generation, and proposes a framework for evaluating bias in LLM-generated code, specifically examining the influence of protected attributes, prompts and web-search capability. We use two metrics: the code bias score (CBS) and the attribute change ratio (ACR), to quantify the prevalence of bias and the degree of influence of different attributes, respectively. In addition, we investigate four lightweight mitigation strategies: Few-Shot, Chain-of-Thought, Few-Shot Chain-of-Thought, and Multi-agent, aimed at mitigating bias in generated code. Our findings reveal that bias remains prevalent across different protected attributes and datasets even after applying mitigation strategies, highlighting the need for more effective approaches to reduce bias in AI-driven code generation systems.

Large language models (LLMs) are increasingly deployed in settings that require nuanced ethical reasoning, yet existing bias evaluations treat model outputs as simply "biased" or "unbiased." This binary framing misses the gradual, context-sensitive way bias actually emerges. We address this gap in two stages: behavioral profiling and mechanistic validation. In the behavioral stage, we introduce the Moral Sensitivity Index (MSI), a metric that quantifies the probability of biased output across a graduated, seven-tier stress test ranging from abstract numerical problems to scenarios rooted in historical and socioeconomic injustice. Evaluating four leading models (Claude 3.5, Qwen 3.5, Llama 3, and Gemini 1.5), we identify distinct behavioral signatures shaped by alignment design: for instance, Gemini 1.5 reaches 72.7% MSI by Tier 5 under socioeconomic framing, while Claude exhibits sharp suppression consistent with identity-based safety training. We then verify these behavioral patterns mechanistically. We select criminal-bias scenarios, which produced the highest MSI scores across models, as probes and apply logit lens, attention analysis, activation patching, and semantic probing to a controlled set of six models spanning three capability tiers: small language models (SLMs), instruction-tuned base models, and reasoning-distilled variants. Circuit-level analysis reveals a U-curve of bias: SLMs exhibit strong criminal bias; scaling to instruction-tuned models eliminates it; reasoning distillation reintroduces bias to SLM-like levels despite identical parameter counts, suggesting distillation compresses reasoning traces in ways that reactivate shallow statistical associations. Critically, the socially loaded cues that drive high MSI scores activate the same bias-driving circuits identified mechanistically, providing cross-stage validation.

Large language models pick up social biases from the data they are trained on and carry those biases into downstream applications, often reinforcing stereotypes around gender, race, religion, disability, age, and socioeconomic status. The standard fixes (retraining on curated data or fine-tuning with human feedback) are expensive, need access to model weights, and risk degrading the model on other tasks. In this paper we take a different route: we debias the model at decoding time, treating bias mitigation as a structured search over candidate tokens without ever touching model weights. A separate Process Reward Model (PRM) acts as a judge, scoring each candidate for both fairness and fluency. We design three schemes of increasing sophistication (Best-of-N selection, Sequential critique-and-revise, and Constitutional self-audit) and evaluate them on four models (GPT-4o-mini, Llama 3.2 3B, Gemma 3 4B, Qwen 2.5 3B) across a 200-prompt bilingual benchmark in English and Urdu covering eight bias categories. Sequential debiasing proves the most effective, raising mean bias scores by up to +0.40 over baseline while preserving (and sometimes improving) fluency. We then extend all three schemes to open-ended generation, where each token is debiased on the fly, and introduce a lightweight Bias Guard gate that fires only on potentially biased words, keeping overhead near 2x for well-calibrated models. A formal overhead metric that separates generator cost from judge cost reveals that Best-of-N is effectively free on the generator side in a native implementation. GPT-4o-mini, included as a strong proprietary anchor, confirms that the framework scales with model capability; the three open-weight models show where current small-scale LLMs still struggle.

Large language models (LLMs) are increasingly used to generate multiple-choice questions (MCQs), where correct answers should ideally be uniformly distributed across options. However, we observe that LLMs exhibit systematic position biases during generation. Through extensive experiments with 10 LLMs and 5 vision-language models (VLMs) on three MCQ generation tasks, we show that these biases are structured, with similar patterns emerging within model families. To investigate the underlying mechanisms, we conduct probing experiments and find that hidden representations in the question stem encode predictive signals of the correct answer position, suggesting that answer position may be implicitly planned during generation. Building on this insight, we apply activation steering to manipulate internal representations and influence answer position. Our results show that steering can partially control positional preferences and substantially shift answer position distributions. Our findings provide a practical framework for studying implicit positional planning in LLMs and highlight the importance of controllable generation for reliable MCQ construction and evaluation.

Prior research has established that instruction-tuned large language models exhibit left-of-center political bias, measured exclusively through abstract political questionnaires. We show that this finding does not generalize to concrete policy decisions. We introduce a dual-instrument methodology grounded in Swiss democratic reality. The Smartvote questionnaire (75 abstract policy questions) is administered to 66 LLMs from 27 model families and compared to 184 elected members of the Swiss National Council, replicating the established leftward convergence (Cohen's d = 3.64, p = 0.0002). Then, novel to this work, 9 flagship LLMs are confronted with 48 real federal referenda (Volksabstimmungen) in four national languages (German, French, Italian, Romansh) under three information conditions, comparing votes to actual outcomes and party recommendations (Parolen). Three findings challenge the prevailing narrative. (1) Abstract questionnaires do not predict concrete behavior: the left-to-right agreement gradient on Smartvote shifts from left-peaked to center-peaked on Volksabstimmungen, where models align most with centrist Die Mitte and FDP rather than leftist SP and Gruene (Wilcoxon p = 0.008). (2) For some models, the language of a political question changes the answer more than the political content does: cross-linguistic consistency ranges from 50% (Mistral) to 98% (GPT-5.4). (3) Two models exhibit systematic change-aversion rather than political bias, voting Nein on 83-94% of referenda regardless of direction (binomial p < 0.0001). What prior work measured as "leftward bias" may not generalize beyond abstract instruments. On concrete policy decisions, LLMs behave less like coalition partners of the left and more like cautious civil servants: centrist, status-quo-favoring, and inconsistent across languages.

Large language models (LLMs) are increasingly used for automated news credibility assessment, yet it remains unclear whether they apply even-handed standards across journalistic genres. We examine whether zero-shot LLMs are more likely to misclassify legitimate entertainment news as fake than legitimate hard news, using a within-dataset design on GossipCop from FakeNewsNet. Across four frontier models, we find a clear but model-specific genre asymmetry: DeepSeek-V3.2 and GPT-5.2 show false-positive-rate gaps of 10.1 and 8.8 percentage points, respectively (both $p < .001$), whereas Claude Opus 4.6 and Gemini 3 Flash show no comparable difference. A style-swap experiment yields only limited and inconsistent changes, suggesting that the asymmetry is not reducible to stylistic register alone. Prompt-based mitigation is likewise possible but not generic: framing the model as an entertainment-news fact-checker reduces false positives for DeepSeek-V3.2 by about 50% without detectable recall loss, but offers little improvement for GPT-5.2. Exploratory qualitative coding further suggests two recurring error patterns in sampled false positives: treating private-life claims as inherently unverifiable and discounting entertainment journalism as an epistemically weaker genre. Taken together, these findings show that aggregate performance metrics can obscure structured false positives within legitimate journalism. We argue that LLM-based credibility assessment may not only evaluate truth claims but also differentially recognize the legitimacy of journalistic genres, and that evaluation should therefore include genre-stratified false-positive analysis alongside overall accuracy.

Large language models (LLMs) are rapidly being integrated into high-stakes public safety systems, including emergency call triage and dispatch decision support, yet their demographic fairness in this context remains largely untested. Here we introduce a cross-lingual audit framework that operationalizes the Police Priority Dispatch System as a five-level ordinal classification task and applies a controlled minimal-pair design to isolate the effect of demographic cues. Across 19,800 model outputs spanning 11 frontier models, 15 scenario pairs, three demographic categories (religious appearance, gender, and race), and two languages (English and Mandarin Chinese), we find that demographic bias emerges systematically when incident severity is ambiguous but largely disappears when the operational priority is clearly determined by call content. Bias magnitude varies by demographic axis, with the largest effects observed for religious appearance, followed by gender and race. Critically, bias does not transfer consistently across languages: gender bias is substantially amplified in Mandarin Chinese, whereas race bias is more pronounced in English, revealing cross-lingual asymmetries that aggregate analyses obscure. In several scenarios, demographic cues produce counter-directional effects, challenging simple stereotype-amplification accounts of model behavior. These findings suggest that bias in LLM-based dispatch is not a fixed property of models alone, but arises from the interaction between demographic signals, contextual ambiguity, and language. Beyond these empirical results, the proposed framework provides a scalable audit infrastructure that enables deploying agencies to evaluate candidate models on jurisdiction-relevant scenarios prior to real-world adoption.

Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment, yet the impact of this compression on model quality remains poorly understood. Existing studies typically compare only two conditions (full-precision vs. a single quantized variant), rely on aggregate bias metrics, and evaluate a single model family, making it impossible to distinguish gradual degradation from threshold-dependent safety failures. We conduct a controlled empirical study of three instruction-tuned models (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items across 5 random seeds, totaling 911,100 inference records. Our results reveal that 3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression, while models' willingness to select "unknown" answers declines by 17.4%. Crucially, these item-level changes are invisible to standard quality metrics: perplexity increases by less than 0.5% at 8-bit and under 3% at 4-bit across all three models, yet 2.5-5.6% of items already develop new biases at 4-bit. These findings demonstrate that aggregate evaluation metrics systematically miss fairness-critical degradation, underscoring the need for quality-aware compression protocols that explicitly test for bias emergence before deployment.