Santiago Grandas, Juan Sebastian Cely-Acosta, Mohit Mendiratta +2cs.CV cs.HC
Beauty assessments from Multimodal Large Language Models (MLLMs) are increasingly popular amongst users, companies, and aestheticians. This raises the question of whether these AI models can accurately reflect human judgments of attractiveness. In a pre- registered exploratory study, we compared the attractiveness ratings of 2,513 human participants to four widely used commercial AI models: Claude, Gemini, GPT, and Grok. Results showed that MLLMs systematically rate faces more favourably and within a narrower range than humans and, at the time of study, do not reproduce human ratings in absolute terms. However, MLLMs exhibit strong correlations with human attractiveness judgments, accurately tracking the rank-ordering of faces. MLLMs may judge faces by different cues than humans; only face age was a predictor of facial attractiveness in both humans and MLLMs, with inconsistent patterns across models for ethnicity and gender. AI models strongly agree with one another, except for Grok, which also showed the lowest agreement with humans. Our findings suggest that while they may be able to approximate rank-orderings of human attractiveness, current off-the-shelf commercial MLLMs systematically overrate the beauty of human faces.
Existing constant-step analysis of stochastic \Scaf{} identifies a leading $O(γ/N)$ stationary mean bias and shows that higher-order bias can persist as the client count increases, but does not identify the first client-independent contribution at coefficient level. For full-participation stochastic \Scaf{} with one-dimensional homogeneous clients, fixed local-step count $H$, and bounded additive gradient noise, we prove, uniformly over $N\ge2$, $$ \begin{aligned} \mathbb{E}_{π_{γ,N,H}}[x]-x^\star ={}& -\frac{f'''(x^\star)σ^2}{4f''(x^\star)^2}\fracγ{N}\\ &- \frac{f'''(x^\star)σ^2}{12f''(x^\star)} \frac{(H-1)(5H-1)}{H}γ^2 +O_H\!\left(\frac{γ^2}{N}+γ^3\right). \end{aligned} $$ Hence client averaging suppresses the leading $O(γ/N)$ bias but does not remove the client-independent $O(γ^2)$ component when its coefficient is nonzero. The mechanism is indirect: although the direct control contribution cancels pathwise in the linear global average, the controls still alter within-round local trajectories and their second moments. Fresh gradient noise and persistent control fluctuations therefore generate local second-moment corrections that nonquadratic curvature converts into stationary mean bias. The coefficient vanishes for quadratic objectives. Numerical experiments are consistent with the predicted coefficient, its persistence as client count increases, and the stated joint remainder. The result is restricted to the one-dimensional homogeneous fixed-$H$ setting.
LLM-based agents are increasingly deployed in persistent social environments, where generated claims can be posted, replied to, remembered, and reused. We study human-directed stereotypes on Moltbook, an open agent-native social platform, asking how agents construct humans as a social category. For this human-target analysis, we introduce an annotation framework with four evaluative dimensions---morality, friendliness, competence, and autonomy---and a second-stage subtype scheme for descriptive \textit{other} attributions. We find that competence dominates human-directed evaluations, while many \textit{other} attributions describe humans as epistemic, cultural, or embodied subjects. We further examine how these human representations appear in human--agent narrative contexts and platform-level circulation. As an auxiliary comparison, we analyze agent-internal community feedback through behavioral host affinity. Rather than reproducing the stable insider--outsider rejection often observed in human online communities, Moltbook feedback patterns are better explained by exposure, author visibility, and content selection. These findings suggest that bias in agent societies should be studied not only as isolated model output, but also as a discourse process.
Monitoring nature restoration at scale is an important but difficult ecological problem. Deep learning methods to analyze satellite image time series (SITS) have been widely used for land surface monitoring. In semi-natural grasslands - the habitat type in focus in this work - restoration outcomes develop gradually, yet satellite observations are influenced by weather, acquisition conditions, and processing artefacts, making it difficult to distinguish genuine restoration signals from unrelated temporal variation. In this work, we examine - to the best of our knowledge, for the first time - whether restoration status can be detected directly from satellite image time series by formulating pasture restoration as a binary deep learning classification problem. We evaluate two common SITS deep learning architectures on different Sentinel-2 image combinations, across 1,397 restored Swedish pastures and find that explicitly modeling intra-year variability and per-pasture normalization increases separability, reaching 0.88 accuracy for the best model. We further investigate our results and perform a targeted bias analysis finding that reliable deployment requires temporally balanced labels and evaluation protocols that explicitly test for year-related confounding. We therefore frame our contribution not as a solved restoration-monitoring system, but as a realistic case study of what works, what fails, and what future studies should control for. Code and models are available at https://github.com/aleksispi/ml-nature-resto.
Professional communication is increasingly mediated by LLMs - but do these models serve all users equally? We show that when prompts contain linguistic features more commonly used by women (hedges, tag questions, collective reference), they systematically elicit shorter, less sophisticated, and less formal responses across three document types and four models. These effects persist after controlling for prompt complexity and feature carry-over. Explicit gender cues like sign-off names are encoded in the same representational space as linguistic dialect - suggesting shared underlying mechanisms - yet linguistic register is far more influential, producing large, consistent effects where names produce none. Our results further reveal that post-hoc mitigation is challenging: because these patterns are culturally embedded and outside conscious control, users cannot easily avoid them through strategic self-presentation, and mechanistic analysis reveals that linguistic features are encoded in early transformer layers and entangled with other features. Our work calls for upstream consideration of the influences of linguistic variation to mitigate disparate impacts of LLM-mediated workplace communication.
Automatic speaking assessment systems are increasingly deployed in high-stakes settings to mark second language (L2) learners' speaking tests, making it critical to show that their scores depend on speaking proficiency rather than irrelevant speaker attributes such as first language (L1) or age. Transformer-based foundation models have improved the accuracy of these L2 speaking graders, but their black-box representations make fairness and interpretability analysis more difficult. Building on prior work that used Concept Activation Vectors (CAVs) to detect bias towards unwanted attributes (`concepts') in feature-based graders, we extend CAV-based analysis to two neural speaking assessment systems: a text-based BERT grader and a speech-and-text multimodal grader based on Whisper. CAVs represent human-interpretable concepts as directions in a model's activation space, allowing us to distinguish between whether a concept is encoded in a model's internal representations and whether it influences the predicted score, the latter quantified using a gradient-based sensitivity metric. Since CAVs rely on linear separability, which is less likely in complex neural embedding spaces, we also investigate whether sparse autoencoders (SAEs) provide cleaner concept directions by learning CAVs in a sparse latent space and mapping them back to activation space. Our analysis shows that concept recoverability depends strongly on the representation and architecture being probed, rather than on the concept alone. Sensitivity to concepts is also architecture-dependent. SAEs make concepts more linearly recoverable, but attenuate the original activation-space sensitivity, especially in low-dimensional layers. These findings highlight the need to distinguish concept recoverability from concept influence when auditing bias in speaking assessment systems.
Cardiac digital twins convert clinical images into physiological measurements through observation operators, yet calibration studies often assume a fixed reference convention. Across four shared-backbone echocardiographic EF front-ends, phase conditioning appears to remove CAMUS baseline bias. Matched-reference analysis rejects this gain: singleplane ground-truth EF error is statistically indistinguishable across models, while single-plane ground-truth EF exceeds CAMUS biplane clinical EF by +6.30 points, explaining nearly all baseline bias. A prespecified EchoNet-Dynamic replication, with released data and our extractor aligned to the apical four-chamber plane, removes baseline overestimation and reverses the CAMUS ranking. We also quantify haemodynamic effects, conformal residual-width budgets, and EF-stratum changes, yielding a Convention-Aware EF Audit protocol that separates genuine observation operator calibration from measurement artefacts. GitHub: EjectionFraction-Bias-in-Cardiac-Digital-Twin.git
Bandit algorithms generate data for downstream inference, but adaptive sampling biases post-bandit sample means. We analyze this bias for stable index algorithms, including UCB1 and its generalizations, and derive sharp leading-order expressions for the sample-mean bias and expected $Z$-statistic. Our characterization reveals the algorithmic origin of bias through a key index-function-dependent quantity, which we term effective exploration rate. For example, under UCB1, the effective exploration rate is of order $\sqrt{\log T}$, and the standardized bias of any arm (that is not uniquely optimal) decays at the extremely slow rate $1/\sqrt{\log T}$. We also show how the choice of the index function affects both regret and bias, which reveals a regret-bias trade-off: more exploratory algorithm reduces bias but increases regret. Our sharp characterization for bias uses a novel empirical fluid approximation of the algorithm's sampling dynamics, which may be of independent interest.
Large language models (LLMs) are increasingly used as stand-ins for human respondents, from opinion polls and simulated survey participants to agent-based social simulations. These uses rest on one assumption: that conditioning a model on who a person is yields answers resembling those of real people from that group. Here we identify and measure benevolence bias, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions. Across 18 widely used models, four social-science datasets (ANES, GSS, WVS, and a cross-cultural prospect-theory replication) and six psychological categories, we find that the bias is a stable model property, not a quirk of any one system: it points the same way across models, grows with model size, and traces to the post-training stage. Prompt language and framing change its size but never its direction, and a "malicious persona" stress test shows a one-sided limit: aligned models struggle to play people who are less kind, less prosocial or more harm-tolerant than average. The issue is thus not only a shifted average, but a narrowed range of people the model can imitate. The bias sits in the middle of the answer distribution rather than its tails, and survives changes in sampling temperature and simple prompted reflection. The encouraging news is that it is easy to diagnose and straightforward to fix: a light-touch contrastive calibration, which needs no retraining and works on black-box APIs, brings all six categories back to the human baseline. Our results give researchers a clear map of where aligned LLMs can already be trusted as human stand-ins, where they need care, and a ready-to-use method for closing the gap.
Yifan Chen, Xiaoou Cheng, Jonathan Niles-Weed +1stat.CO cs.LG math.PR stat.ML
Unadjusted samplers such as unadjusted Hamiltonian Monte Carlo and underdamped Langevin are well-known to be biased. Metropolis--Hastings adjustment has been conventionally incorporated into Hamiltonian Monte Carlo to eliminate the bias. However, this adjustment can significantly increase the iteration complexity due to the small step size required for reasonable Metropolis acceptance rates. In this work, we extend the \emph{delocalization of bias} phenomenon, previously established for the overdamped Langevin algorithm, to these two unadjusted algorithms. We show that to control the $W_2$ bias of any $K$-dimensional marginal of a high-dimensional distribution, $O(\sqrt{K})$ integration steps suffice up to $\log d$ terms, assuming either weak or sparse interactions among variables. The discrete-time integrators here introduce technical difficulties beyond those of the overdamped setting, which we address through a broadly applicable matrix-polynomial framework that characterizes their propagators. Our result for the underdamped Langevin algorithm is valid for all large friction parameters, implying that the Leimkuhler-Matthews integrator for the overdamped Langevin dynamics also exhibits delocalization of bias.
LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.
Despoina Giarimpampa, Roland Meier, Tegawendé F. Bissyandé +2cs.CY cs.AI
Expert surveys are widely used in security research to study practitioner workows and decision-making, yet recruiting domain experts - especially in Security Operations Centres (SOCs), where analysts face high workload, burnout and confidentiality constraints - is difficult and often results in small samples. Large language models (LLMs) oer an appealing alternative by generating synthetic responses at scale, but little guidance exists on when such surrogate participants are reliable. We present a methodological framework for evaluating LLMs as substitutes or supplements to expert survey respondents. Using responses from SOC professionals, we compare persona-based and aggregate LLM-generated answers across multiple models and prompting settings. We measure stability, inter-model agreement and alignment with human responses. Our results show that although LLMs produce internally consistent answers, they systematically diverge from experts, exhibiting reduced variance, central tendency bias and homogenised opinions. This work contributes methodological evidence and practical guidance to the security research community on the appropriate use and limitations of LLM-generated survey responses. We conclude that LLMs are useful for piloting and hypothesis generation but not for replacing expert elicitation, and we discuss implications for researchers using LLM-augmented surveys.
Xiangyu Ma, Mengmi Zhang, Shannon Ang +1cs.CL econ.GN
Large-language models have proven to be remarkable if inconsistent parrots of public attitudes and opinions. The extent to which LLMs are able to produce reasonable approximations of cultural taste remains an open empirical question that becomes more urgent by the day, with market research companies already offering provisional `synthetic' survey panels and the contamination of standard survey data from LLM-generated responses. In this study, we build on past work on silicon sampling by extending considerations of its algorithmic fidelity and alignment to the domain of cultural consumption. We use large-language models from OpenAI, Anthropic, and DeepSeek to each produce 277,470 (30x9249) silicon surrogates of survey respondents from the Survey of Public Participation in the Arts (SPPA). We find these silicon surrogates' tastes to be highly stylized facsimiles of human tastes. (1) Silicon samples have a systematic postive-bias for liking, resulting in inflated ecological estimates of tastes. The individual-level bias of silicon samples are not well-explained by the WEIRD-bias often discussed in the literature. (2) The complex relationality in real taste structures is completely lost among silicon samples. (3) Finally, very little of the known cultural alignment between tastes and social space are preserved. Silicon samples attenuate age-taste associations, resurrect anachronistic class-taste associations, caricaturize gender- and race-taste associations.
This study investigates sentiment polarity biases, specifically, differences in how accurately AI models classify positive versus negative reviews across languages and model architectures. Large language models show a negative bias in French and are more accurate on negative reviews, while encoder models exhibit positive bias in Japanese, missing negative reviews that use indirect criticism. These language-specific polarity biases have implications in both social and business domains deploying multilingual sentiment analysis systems.
The global use of artificial intelligence has increased interest in assessing the ability to generate culturally localized content, including stories. Cultural localization in stories often occurs through either templated localization -- the use of cultural markers (e.g., names, locations) in a generic narrative -- or holistic localization -- the variation of plots, values, and themes, in addition to cultural markers. We propose a method to measure the degree to which content was generated through templated localization. Specifically, we identify the lexical tokens that distinguish stories across nationalities and measure the similarity of the narratives that remain after removing them. In stories generated by five models on 125 topics for 193 nationalities, our method is able to detect that only a small subset (9-17%) of the vocabulary accounts for the variation across nationalities and that the narratives that remain after removing them contain repeated multi-word sequences, suggesting the presence of a shared culturally-agnostic narrative template. Finally, we characterize the cultural markers for their stereotypicality and offensiveness, finding that markers from 19 countries, mostly located in the Global South, are on average offensive.
Current research primarily focuses on model performance, while comparatively less attention has been devoted to uncertainty estimation, particularly in settings where LLMs are increasingly used to generate annotated data. We introduce a framework combining conformal prediction with Collaborative Filtering-style annotators' representation to model LLM behavior in relation to human annotators and to analyze patterns of agreement and disagreement. Using Non-Conformity Scores, we introduce the Ghost Prediction metric and the Ghost Annotator representation to quantify cases in which model predictions diverge from all available human annotations. We compute cosine similarity measures to explore differences in model behavior across sociodemographic axes. We evaluated four LLMs of different size and families across four content moderation datasets. Our finding shows that while we find that all models uncertainty increases with annotator disagreement, larger models tend to be more confident in the classification of texts that are not aligned with any human annotation. Finally, the Ghost Annotator framework reveals a consistent and robust pattern of demographic misalignment, suggesting a structural bias likely rooted in pretraining corpora.
Francisco Valentini, Edgar Altszyler, Martin Fajcikcs.IR cs.AI cs.CL
Neural retrievers are trained to estimate query-document relevance from annotated query-document pairs. Yet annotation protocols may not purely reflect relevance: they select only a subset of documents for labeling, and this selection can favor certain document types over others. We investigate whether supervised bi-encoder retrievers implicitly learn a document-level relevance prior: a query-independent signal encoded in their representation space as a side effect of training on annotated data. We estimate this prior by training simple classifiers on frozen document embeddings and evaluate three state-of-the-art retrievers across multiple IR benchmarks. We find that supervised neural retrievers encode relevance priors that generalize to unseen documents and are consistent across models. These priors create a findability gap: documents with lower prior are systematically harder to retrieve, even when genuinely relevant. This effect appears in supervised dense retrievers but is weaker and less consistent in BM25, and it persists under controlled matched-document comparisons. Using LLM-based explanations, we find that judged-relevant documents tend to be comprehensive, self-contained summaries of mainstream topics, while niche, fragmentary, or highly technical content is often left unjudged. Retrievers internalize this bias, ranking documents with these favored features higher than documents that lack them, independently of their actual relevance. Our findings expose a structural limitation of supervised retrieval: models trained on annotated data do not just learn relevance, but also the implicit document preferences in their training data.
Promotional language has been increasingly used to aid the communication of innovative ideas in science. Yet, less is known about its role in the context of technological innovation. Here, we use a validated and domain-diagnosed lexicon of 135 promotional words to study the association between promotional language and patent evaluation outcomes among 2.7 million USPTO patent applications. Our large-scale study reveals three unexpected findings. First, in contrast to scientific evaluation, we find that a higher frequency of promotional words is negatively associated with the probability of an application being (i) granted a patent, (ii) transferred ownership, and (iii) successfully appealed. This promotional penalty holds even after accounting for a range of confounding factors and is largely robust across different technological areas. Among matched samples, the difference in the success rate between the lowest and highest promotional density quintile is 5.5, 5.9, and 5.3 percentage points for patentability, transferability, and rejection reversal. Second, contrary to institutional skepticism, we show that promotional language is not a mask of weak technology, but objectively reflects the degree of combinatorial novelty and future citation impact. Third, digging into the mechanisms, we find that the tolerance to promotional framing is strongly moderated by human factors, with men and experienced examiners showing a higher acceptance of promotional narratives than women and novice examiners. By revealing an emerging paradox in the patent system, our study offers theoretical and practical implications for improving patent evaluation through more objective scrutiny of linguistic patterns in patent filings.