Implicit Bias

Assessing progress toward the Sustainable Development Goals (SDGs) requires multi-step reasoning over visual cues, contextual knowledge, and development indicators, where incomplete evidence use and imperfect evidence integration can introduce hidden prediction biases. Real-world SDG monitoring further spans both qualitative judgments and quantitative estimation. However, existing benchmarks typically evaluate these aspects in isolation, obscuring systematic biases that emerge when models substitute priors for evidence. To address this gap, we propose SDGBiasBench, a large-scale benchmark suite for SDG-oriented vision-language reasoning. Spanning 500k expert-involved multiple-choice questions and 50k regression tasks, the benchmark enables comprehensive assessment of both decision-level and estimation-level bias in Vision--Language Models (VLMs). Evaluations on SDGBiasBench reveal an intrinsic SDG bias in current VLMs, where predictions are frequently driven by SDG specific priors rather than reliable multi-modal cues. To mitigate such bias, we propose CADE (Contrastive Adaptive Debias Ensemble), a training-free, plug-and-play method that leverages modality-specific answer priors. CADE yields significant gains on the proposed benchmark, improving multiple-choice accuracy by up to 25% and reducing regression MAE by up to 12 points across multiple VLMs. We hope our work can foster the development of more fair and reliable AI systems for sustainable development.

Statistical decision algorithms are increasingly deployed in domains where ground-truth labels are hard to obtain, such as hiring, university admissions, and content moderation. In these settings, models are typically trained on historical human evaluations -- for example, using past hiring decisions as a proxy for true applicant quality. However, if past evaluations unjustly favor certain groups, models trained on these labels may inherit those biases. To address this problem, we propose basing predictions on rubric embeddings, a representation framework that replaces standard black-box embeddings with features derived from expert-defined criteria that align with the underlying construct of interest. By anchoring predictions to semantically meaningful dimensions, this approach guards against biased proxy signals. We provide both theoretical and empirical evidence that rubric embeddings mitigate label bias under plausible conditions. Empirically, we evaluate our method on a novel dataset of applications to a large master's program. We find that models trained on rubric embeddings reduce group disparities while improving measures of cohort quality. Our results suggest that basing predictions on interpretable, domain-grounded representations offers a practical approach to learning in the presence of biased labels.

Functional bilevel methods estimate a lower-level function and plug it into a hypergradient, but this plug-in gradient can retain first-order bias when the lower-level problem is learned nonparametrically. To remove this bias, we develop a semiparametric debiasing theory for population bilevel gradients based on the efficient influence function. This perspective leads to a cross-fitted orthogonal hypergradient estimator for which we establish asymptotic normality together with uniform control over the outer parameter. Under quadratic losses, the estimator reduces to a simple doubly robust score based on conditional mean nuisances. On synthetic bilevel benchmarks with known ground truth, the method tracks the oracle efficient-gradient benchmark and improves over plug-in functional hypergradients and regularized kernel bilevel baselines.

Unsupervised feature selection is commonly formulated as a multiobjective optimisation problem that jointly optimises subset quality and subset size. Yet the behaviour of this formulation depends critically on the choice of evaluation objective, the direction of subset-size regularisation, and the initialisation strategy. We study these factors in a controlled setting using a synthetic dataset with known informative, redundant, and irrelevant feature types. Six formulations are compared by combining three evaluation objectives: accuracy, silhouette score, and PCA reconstruction loss with subset-size minimisation or maximisation. The results show that formulation strongly affects both search dynamics and the quality of the resulting Pareto front. Silhouette-based formulations exhibit a strong bias toward trivial low-cardinality solutions and remain weak proxies for predictive performance. In contrast, the proposed PCA loss objective produces compact subsets with test accuracy comparable to subsets obtained by directly optimising supervised accuracy. Overall, the study shows that objective design is central to effective multiobjective unsupervised feature selection.

We study the effect of different persona on \textbf{sycophancy}: model's agreement with users even when the user is incorrect. The standard mitigation, Contrastive Activation Addition (CAA), derives a steering direction from labelled pairs of sycophantic and honest responses. This study evaluates whether off-the-shelf persona steering vectors, originally developed for general role-playing and not trained on sycophancy data, can serve as an alternative. In two instruction-tuned models, steering toward personas characterised by doubt or scrutiny reduces sycophancy to approximately $68\%$ and $98\%$ of CAA's effect, and, unlike CAA, maintains accuracy when the user is correct. The effect is also asymmetric: steering toward agreeable personas does not produce a mirror increase in sycophancy. Geometrically, the persona vector is largely independent of the direction of sycophancy in activation space. Collectively, these findings suggest that sycophancy is better understood as a persona-level property rather than a single steerable direction. We release our code here: https://anonymous.4open.science/r/Sycophancy-Steering-9DF0/.

As Artificial Intelligence (AI) literacy initiatives expand in K-12 settings, understanding how gender shapes student baseline perceptions, tool-use, and responsiveness to interventions is essential for equitable curriculum design. This study examines gender differences in AI literacy, safety awareness, and STEM career aspirations among Australian secondary students (Years 7, 8, and 10; N(pre) = 199, n(post) = 136) from two co-educational government schools who participated in a one-day AI literacy workshop. Using statistical regression methods controlling for year level and school, we found that pre-workshop, male students reported significantly higher STEM career interest across all three domains (AI, computer science, and engineering), while female students were significantly more likely to use AI for schoolwork and to seek advice from AI tools. Gender-differentiated patterns also emerged in deepfake behaviours: males were significantly more likely to have created or shared deepfake content. Both genders improved in AI knowledge post-intervention, yet females showed a richer profile of gains: wider conceptual understanding, greater confidence, and meaningful increases in AI and CS career interest that partially narrowed the gender STEM gap. These findings highlight the need for gender-responsive AI curricula, particularly deepfake safety education for male students, and demonstrate that even single-day workshops can narrow gender gaps in STEM aspirations and AI confidence.

Large language models (LLMs) are increasingly deployed in socially sensitive settings despite substantial documentation that they encode gender biases. Chain-of-Thought (CoT) prompting has been proposed as a bias-mitigation approach. However, existing evaluations primarily focus on changes in LLM benchmark performance, providing limited insight into whether apparent bias reductions reflect meaningful changes in a model's internal mechanisms. In this work, we investigate how CoT prompting affects gender bias in LLMs, combining benchmark-based evaluation with mechanistic interpretability techniques and reasoning chain failure analysis. Our results confirm a stereotypical bias present in LLM outputs across benchmarks, showing that CoT prompting does not consistently reduce the bias gap. Mechanistic analyses reveal that although CoT balances biased behavior in certain attention head clusters, gender bias remains embedded in hidden representations, indicating only superficial mitigation. Inspection of reasoning chains further suggests that these improvements stem from memorization and familiarity with the dataset rather than genuine understanding of bias.

This work investigates the ``small-vs-large gap'', where repeating on fewer samples can lead to compute saving during training compared to using a larger dataset. This is observed across algorithmic tasks, architectures and optimizers and cannot be explained using prior theory. We argue that the speedup comes from appropriate layer-wise growth enabled by sampling biases, which is more pronounced when the dataset size is smaller. We provide both theoretical analysis and empirical evidence from various interventions. Our results suggest that using a smaller dataset with more repetitions is not just a fallback strategy under data scarcity, but can be proactively leveraged as a favorable inductive biases for optimization, particularly in reasoning tasks.

Social roles shape expectations, priorities, and judgments, yet it remains unclear how large language models (LLMs) associate occupational identities with broader cultural value patterns. Prior work used nationality-based cultural prompting to study how LLM responses to value-survey questions align with human cultural benchmarks. In this paper, we extend that framework by replacing cultural prompting with occupational prompting to examine how professional-role cues influence value-survey responses in open-weight LLMs. Using a survey-grounded evaluation pipeline based on questions from the Integrated Values Surveys, we project model responses into the two-dimensional Inglehart--Welzel cultural space. We prompt open-weight LLMs to answer questions under occupational identities such as accountant, teacher, engineer, and nurse, and then analyze how these occupation-conditioned responses are positioned on the cultural map. Our results show that when open-weight LLMs are prompted with occupations rather than national identities, their responses remain within a broadly Western-leaning region of the cultural map. However, different occupations introduce shifts within this region, producing distinct occupational skews. This indicates that occupational prompts are not treated as neutral role labels, but instead elicit structured value patterns. These findings extend survey-based evaluation of cultural bias beyond nationality-based prompting and provide a framework for studying how occupational personas shape value expression in LLMs.

Estimating forest aboveground biomass (AGB) from Earth observation combines two structurally incompatible label sources: spaceborne lidar provides canopy structure at millions of locations but no biomass estimate, and ground-based plots provide biomass at thousands of biased locations but no metrics of structure. No single training sample carries labels for all target variables, plot labels are missing not at random (MNAR), and biomass is linked to the structural variables by known but biome-specific allometric laws. We formalise this as multi-task dense regression under heterogeneous disjoint partial supervision with MNAR labels and inter-task physical constraints, and propose StruMPL to address it jointly. A shared encoder feeds per-variable regression, imputation, and propensity heads for spatial MNAR correction, and a learnable physics module that evaluates the inter-task constraint on the model's own predictions at every pixel. The supervised loss uses an Augmented IPW (AIPW) pseudo-outcome with stop-gradients on the propensity and on the imputation baseline; we show analytically and empirically that both are necessary for joint optimisation to recover IPW-weighted stationary points while keeping the loss bounded. On two ecologically distinct biomes, StruMPL outperforms ablation variants and the closest published method on AGB RMSE and bias, with a stratified analysis showing AIPW reduces high-AGB bias by ~54%.

Strategic classification(SC) studies the interaction between decision models and agents who strategically manipulate their features for favorable outcomes. Existing SC frameworks typically rely on the idealized assumption that agents are strictly rational. However, evidence from behavioral economics and psychology consistently shows that real-world decision-making is often shaped by cognitive biases, deviating from pure rationality. To formalize this limitation, we identify and define a new problem setting, termed the behaviorally realistic strategic classification problem, where agents' strategic manipulations deviate from full rationality due to psychological biases. Motivated by the identified limitation, we propose the Prospect-Guided Strategic Framework (Pro-SF) to address the problem, a principled framework grounded in prospect theory to model and learn under behaviorally realistic strategic responses. Specifically, to capture behaviorally realistic strategic manipulations, our framework reformulates the Stackelberg-style interaction between agents and the decision-maker by incorporating three key mechanisms inspired by prospect theory, including the asymmetry between benefits and costs, different subjective reference points, and non-rational probability distortion. Experiments on synthetic and real-world datasets establish Pro-SF as a behaviorally grounded approach to strategic classification, bridging machine learning and behavioral economics for more reliable deployment in the real world.

We study the max-margin solutions reached by mirror flow in deep neural networks with homogeneous activation functions. Extending classical results on gradient flow, we derive a novel balance equation for mirror flow from convex duality, enabling a characterization of the horizon function governing the induced margin. We further establish max-margin characterizations together with convergence rates and norm growth estimates. Finally, we support our theory through experiments on synthetic datasets and standard vision tasks. Concretely, we show that: (1) distinct non-homogeneous mirror maps can induce the same max-margin solution; (2) convergence can be extremely slow, including exponentially slow regimes; and (3) although all considered mirror maps exhibit feature learning, they can produce markedly different representations, ranging from sparse to dense neuron activations. Together, these results provide a unified perspective on sparse and dense feature learning in homogeneous neural networks, highlighting how mirror maps shape both optimization dynamics and the geometry of the learned classifiers.

Recent advances in Video Large Language Models (Video-LLMs) have greatly expanded multimodal reasoning capabilities. However, the massive number of visual tokens extracted from long video sequences incurs prohibitive computational costs, limiting their deployment in real-world scenarios. Existing training-free token compression methods select tokens based on attention magnitude as a proxy for semantic importance, but often overlook positional bias and rely only on short-term temporal locality, leading to redundant spatio-temporal coverage and inefficient token usage. We present DynaTok, a training-free, temporally adaptive and bias-aware token compression framework that allocates token budgets across both temporal and spatial dimensions. Through a lightweight exponential moving average (EMA) memory, the Temporal Budget Allocation (TBA) module dynamically assigns fewer tokens to redundant frames and more to novel frames, capturing long-term temporal variation. The Spatial Budget Allocation (SBA) module complements this by selecting spatially diverse and semantically important features using activation-based attention maps, while leveraging a spatial memory to reduce redundancy from previously selected regions and mitigate positional bias. DynaTok integrates seamlessly with existing Video-LLMs such as LLaVA-OneVision and LLaVA-Video without retraining, and effectively preserves semantic coverage under aggressive compression. Experiments on four representative VideoQA benchmarks-MVBench, LongVideoBench, MLVU, and VideoMME-show that DynaTok retains over 95% of baseline accuracy even with a 90% token reduction, surpassing recent training-free approaches. These results demonstrate that DynaTok provides a principled foundation for efficient and robust video reasoning, paving the way toward real-time streaming video understanding with future Video-LLMs.

Continual Test-Time Adaptation (CTTA) aims to empower perception systems to handle dynamic distribution shifts encountered after deployment. Existing methods predominantly follow a backward-alignment paradigm, which rigidly aligns incoming data with supervisory surrogates derived from the source domain. Consequently, they struggle with unreliable supervision and evolving distribution shifts. To overcome these limitations, we introduce a novel forward-facilitation paradigm through a method termed Dynamic Style Bridging. Prior to deployment, we construct a compact knowledge base of generated class exemplars. During test time, to mitigate inherent generative bias and adapt these proxies to incoming data, we propose a multi-level bridging mechanism. This mechanism dynamically injects the proxies with incoming data styles at the input, statistical, and representation levels, while preserving the original semantics of the proxies. These high-fidelity proxies are then used to provide reliable, on-demand supervisory signals, enabling stable adaptation under continual shifts. Extensive experiments across standard CTTA benchmarks demonstrate that our method achieves consistent and substantial improvements over recent state-of-the-art approaches. Code is available at \href{https://github.com/z1358/DAS}.

Interpreting the decisions of deep image classifiers remains challenging, particularly in black-box settings where model internals are inaccessible. We introduce OCCAM, a framework for open-set causal concept explanation and ontology induction in vision models. OCCAM discovers visual concepts in an open-set manner, localizes them via text-guided segmentation, and performs object-level interventions by removing concepts to measure changes in class confidence, estimating each concept's causal contribution. Beyond local explanations, OCCAM aggregates interventional evidence across a dataset to induce a structured concept ontology that captures how classifiers globally organize visual concepts. Reasoning over this ontology reveals consistent dependencies between concepts, exposes latent causal relations, and uncovers systematic model biases. Experiments on Broden and ImageNet-S across multiple classifiers show that OCCAM improves explanation quality in open-set black-box settings while providing richer global insight than per-image attribution methods.

Large language models (LLMs) exhibit systematic political biases in voter simulations, but their underlying mechanisms and cross-lingual generalizations remain poorly understood. We introduce Dynamo-K, a census-grounded simulation framework evaluating Korean-language LLM political behavior across four models on six Korean elections (2017-2025). Using this framework, we identify three systematic failure modes: (1) progressive bias in moderate agents, where explicit mitigation reduces Mean Absolute Error (MAE) by 5.2 times; (2) model-dependent third-party salience collapse, distinguishing between salience failure and decision bias; and (3) regional polarization collapse, where models bidirectionally under-predict historical party strongholds. To address these failures, we demonstrate that scenario reframing recovers 62% of 2017 MAE by restoring third-party visibility. Furthermore, we introduce a learned reweighting adapter that successfully calibrates opposing-valence models without relying on candidate names at train or test time. Validating our diagnostic framework, Dynamo-K accurately predicts 3/3 presidential winners - including a 2.1%p MAE on the highly contested 0.73%p-margin 2022 race - and correctly identifies the dominant party in a held-out local election. The pipeline is open-source and provides a scalable, cost-effective method for diagnosing LLM political behavior.

Vision-Language Models (VLMs) excel at multimodal reasoning, yet it remains unclear whether their answers are grounded in visual evidence or driven by learned language and world priors. Counting provides a precise testbed: when visual evidence conflicts with canonical object knowledge, a model must rely on the image rather than a prototypical count. We introduce CounterCount, a diagnostic framework for counterfactual counting in VLMs, consisting of paired factual and counterfactual images with edited count-relevant attributes, verified answers, and localized evidence annotations. Evaluating recent VLMs, we find strong performance on factual images but consistent degradation under counterfactual attribute changes, indicating reliance on object-level priors even when contradictory visual evidence is present. Using localized annotations, we show that these failures are not solely due to missing or ambiguous visual evidence, but to models underweighting attention to count-relevant visual tokens. We introduce a unified inference-time attention modulation strategy that reweights selected visual tokens, improving counterfactual counting accuracy by up to 8% across multiple VLMs. Overall, CounterCount exposes prior-driven counting failures and provides diagnostic insights for designing future VLMs.

Power differences shape human communication through well documented socio cognitive effects, including language coordination, pronoun usage, authority bias, and harmful compliance. We examine whether large language models (LLMs) exhibit similar behaviors when assigned high or low status personas. Using personas from diverse professions, we simulate multi turn, power asymmetric dialogues (e.g., principal teacher, justice lawyer) and measure (i) language coordination, (ii) pronoun usage, (iii) persuasion success, and (iv) compliance with unsafe requests. Our results show that LLMs show key socio-cognitive effects of power, albeit with nuances and variability, linking simulated interactions to both desirable and unsafe behaviors.

Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support and medical documentation. However, the robustness of these models against subtle linguistic variations, specifically stigmatizing language (SL) commonly found in human-authored clinical notes, remains critically under-explored. In this work, we investigate whether frontier LLMs inherit and propagate this human bias when processing clinical text. We systematically evaluate nine frontier LLMs across four stigmatized medical conditions, utilizing clinical vignettes injected with varying intensities and phenotypes of SL (doubt, blame, and maligning). Our results demonstrate that all evaluated models exhibit substantial bias, with clinical decision-making significantly skewed towards less aggressive patient management. Notably, we observe a high sensitivity to linguistic framing, where a single SL sentence is sufficient to alter model outputs, revealing a clear dose-response relationship. Furthermore, we evaluate standard prompt-based mitigation strategies, including Chain-of-Thought (CoT) reasoning and model self-debiasing. These approaches show limited efficacy; models struggle to explicitly identify SL while remaining implicitly influenced by it. Our findings expose a critical vulnerability in current LLMs regarding fairness and robustness in clinical NLP, underscoring the need for rigorous algorithmic guardrails to prevent the automation of health disparities.

Deployed language models are evaluated in a non-stationary environment: model versions, retrieval layers, safety systems, and real-world inputs all change over time. Static bias benchmarks remain useful, but they do not show how models frame newly emerging events for different prompted audiences. We introduce GPF-LIVENEWS, a streaming evaluation protocol and benchmark snapshot for auditing group-conditioned framing in open-ended LLM outputs. The protocol expands fresh BBC/Reuters news anchors across 42 identity labels and seven prompt families, then evaluates response bundles using semantic-sensitivity and sentiment-disparity signals. In a pilot over 12 monitoring runs and 23 hosted models, Policy/Action prompts produce the strongest semantic movement, while sentiment variation is flatter across dimensions and prompt families. The released artifact includes article metadata, prompt templates, instantiated prompts, model-output metadata, score tables, documentation, and reproduction scripts. We interpret all scores as observed-window audit signals for human review, not as permanent fairness rankings or direct proof of harmful bias.

LLM agents exhibit a consistent tendency to over-call, invoking tools even in situations where none is needed. On the When2Call benchmark, six models from three families show high call accuracy but much lower no-call accuracy, leaving overall accuracy in the 55%-70% range. We trace this to an Intrinsic Bias Hypothesis (IBH): the call/no-call decision mapping carries an activation-independent call offset, so the model favors call even at activation parity. Using Sparse Autoencoders (SAEs), we recover behavior-aligned feature bases for the call/no_call decision, reduce them to a signed activation margin, and estimate the offset directly. Across all six models, the model is decision-neutral only when no_call activation outweighs call activation, consistent with IBH. We then causally test IBH with Adaptive Margin-Calibrated Steering (AMCS), a closed-form counter-bias shift along SAE decoder directions. Cancelling the diagnosed offset mitigates over-calling and improves overall accuracy with a negligible drop in call accuracy. Our work recasts over-calling from an empirical phenomenon into a mechanistic object amenable to causal correction. Code is available at https://github.com/SKURA502/agent-sae/.

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 to simulate human opinions and survey responses, but their ability to reproduce population responses across cultures remains limited. Existing persona-based prompting methods typically rely on sociodemographic or personality traits, which are only indirect proxies for the values that shape human responses. We propose a value-based persona construction method that derives textual descriptors from survey responses capturing core cultural dimensions. By sampling value profiles from target populations and aggregating LLM responses across personas, we obtain population-level predictions grounded in observed value distributions. We further introduce a calibration procedure that improves response diversity while preserving estimated opinions. We show that our approach reduces prediction error across countries, with the largest improvements observed in underrepresented populations. This substantially narrows the performance gap between countries aligned with dominant LLM priors and those that are less represented in training data, while also yielding response distributions that closely match human diversity.

Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.

We developed a tool for detecting domain shifts, namely subtle differences in the probability distributions of datasets. We identify these shifts using an algorithm designed to detect localised density anomalies in high-dimensional feature spaces. If an anomaly is present, we then identify the feature subspace in which the anomaly is most pronounced. This allows us to trace the domain shift to a small set of features, making the shift interpretable. Moreover, we provide a protocol for compensating domain shifts by extracting, from two unlabelled datasets, subsets of samples with no detectable residual distributional difference. We validate the framework on controlled 20-dimensional benchmarks with known ground truth, recovering both broad and localized shifts together with their supporting feature subspaces. We then apply it to healthy electrocardiogram (ECG) recordings represented by 782 features. In age- and sex-matched cohort comparisons differing in measurement-device composition, the method detects device-induced shifts, extracts representative subsets enriched in the imbalanced device components, and identifies ECG features associated with the acquisition contrast. These results suggest that density-shift detection and subspace attribution provide a practical framework for uncovering hidden cohort biases before downstream modelling.

Through an analysis of sequences generated by open-weight large language models (LLMs), we demonstrate that LLMs are heavily biased toward English. While continual pre-training is commonly used to adapt LLMs to a target language, we show that it does not offer a cost advantage over training from scratch, even for improving cultural understanding in the target language. These findings suggest that dedicated per-language investment may become increasingly important for future LLM development, rather than relying primarily on the expansion of English-centric resources.

Financial, social, and political factors often prevent the interests of the owners of ML systems and services and their users from being perfectly aligned. ML systems often produce biased information that can influence users to make decisions that are not in their best interest. Current solution approaches require ML systems to implement protocols to mitigate their biases. However, ML system owners usually do not have any incentive to implement these protocols and often argue that it limits their freedom of expression or business. We believe that a successful solution to this problem must recognize the conflict of interest between the ML systems and their users, and use this information to protect users against information that adversely influences their decisions while allowing users to safely benefit from these systems. To this end, we propose a game-theoretic framework that models the interaction between ML systems and users with conflicts of interest. We present scalable algorithms with theoretical guarantees that maximize the amount of desired information and actions and minimize the amount of biased and manipulative actions in interaction with ML systems.

Large-scale facial datasets like CelebA are widely used in computer vision, yet the cultural biases embedded in their labels remain underexplored. Fairness research has distinguished representational from allocational harms, but audits of computer vision datasets have mostly examined categorical labels, leaving open how such harms appear in learned features and model attention. This paper examines CelebA at three levels: dataset structure, learned feature weights, and spatial attention, focusing on how gendered double standards of ageing and beauty are encoded in the data and reproduced in model behaviour. First, hierarchical clustering of 202,599 images shows that the 39 attributes organise into latent trait bundles aligned with cultural archetypes: performative femininity (youth, makeup, adornment) and professional masculinity (ageing, facial hair, formal attire). Female faces, though more often rated attractive overall, incur steep penalties when assigned to ageing or masculine-coded clusters. Second, XGBoost with SHAP analysis reveal gender-specific effects, such as adiposity reducing attractiveness only for females. Third, Grad-CAM finds that predictions for female and younger male subgroups concentrate on mid-face cues, whereas predictions for older males drift toward peripheral cues such as hair and clothing. Older males attain the highest accuracy but the lowest average precision, indicating categorical exclusion of groups outside the dataset's evaluative templates. Cultural double standards thus pass from media representation into dataset labels, feature weights, and model attention, producing two representational harms: hyper-scrutiny of women under a narrow evaluative template, and exclusion of older men from the scheme entirely. Fairness metrics focused on performance disparities mask both, underscoring the need to address representational harm in fairness research.

EEG-based visual decoding aims to establish a mapping between neural signals and visual semantics. However, it remains constrained by the dual challenges of severe information granularity mismatch and the low signal-to-noise ratio (SNR) of EEG signals. Existing approaches typically treat static visual features, ignoring the dynamic selectivity of human vision and the frequency specificity of neural oscillations. To bridge this gap, we propose CAIA, a Cognitive-guided Adaptive blurring with Information-Constrained Alignment framework for Neural-Visual decoding. On the visual side, it simulates selective attention to adaptively reduce redundancy. Meanwhile, on the EEG side, it leverages neural oscillation priors and the information bottleneck mechanism to enhance SNR. Specifically, we devise a cognitive-dynamics-based adaptive blurring mechanism that dynamically integrates center-biased and saliency-guided visual cues via cross-modal attention. Furthermore, we introduce a distribution-aware boundary calibration loss to robustly rectify alignment bias caused by outlier samples. Moreover, a cognitively-guided information-screening method is proposed to select task-relevant EEG oscillations. Extensive experiments demonstrate that CAIA improves both subject-dependent and subject-independent average Top-1 and Top-5 accuracy in zero-shot brain-to-image retrieval, significantly outperforming prior methods. Our work validates that optimizing visual information density to match neural granularity offers a more interpretable and robust pathway for neural decoding.

Large language models are often described as sycophantic, in the sense that they appear to flatter users or mirror their beliefs. We argue that this label is conceptually misleading: sycophancy implies motives and strategic intent, which LLMs do not possess. Their behaviour is better understood as complacency, a structural tendency to agree with user input because training data, reward signals and design favour agreement and reinforcement over correction. We argue that this distinction matters. Whether developers act sycophantically or not, models themselves never are sycophants; they can only be made more or less complacent. This reframing locates agency in developers and institutions, not in the model. Because complacent models reinforce users' prior beliefs, we argue that AI literacy educational approaches should particularly focus on strategies to counter confirmation bias.

As the misuse of AI-generated images grows, generalizable image detection techniques are urgently needed. Recent state-of-the-art (SOTA) methods adopt aligned training datasets to reduce content, size, and format biases, empowering models to capture robust forgery cues. A common strategy is to employ reconstruction techniques, e.g., VAE and DDIM, which show remarkable results in diffusion-based methods. However, such reconstruction-based approaches typically introduce limited and homogeneous artifacts, which cannot fully capture diverse generative patterns, such as GAN-based methods. To complement reconstruction-based fake images with aligned yet diverse artifact patterns, we propose a GAN-based upsampling approach that mimics GAN-generated fake patterns while preserving content, size, and format alignment. This naturally results in two aligned but distinct types of fake images. However, due to the domain shift between reconstruction-based and upsampling-based fake images, direct mixed training causes suboptimal results, where one domain disrupts feature learning of the other. Accordingly, we propose a Separate Expert Fusion (SEF) framework to extract complementary artifact information and reduce inter-domain interference. We first train domain-specific experts via LoRA adaptation on a frozen foundational model, then conduct decoupled fusion with a gating network to adaptively combine expert features while retaining their specialized knowledge. Rather than merely benefiting GAN-generated image detection, this design introduces diverse and complementary artifact patterns that enable SEF to learn a more robust decision boundary and improve generalization across broader generative methods. Extensive experiments demonstrate that our method yields strong results across 13 diverse benchmarks. Codes are released at: https://github.com/liyih/SEF_AIGC_detection.

As generative AI models such as large language models (LLMs) become more pervasive, ensuring the safety, robustness, and overall trustworthiness of these systems is paramount. However, AI is currently facing a reproducibility crisis driven by unreliable evaluations and unrepeatable experimental results. While human raters are often used to assess models for utility and safety, they introduce divergent biases and subjective opinions into their annotations. Overcoming this variance is exceptionally challenging because very little data exists to study how experimental repeatability actually improves as the annotator pool grows. Standard evaluation practices typically rely on a small number of annotations per item (often 3 to 5) and lack the persistent rater identifiers necessary to model individual variance across items. In this work, we introduce a multi-level bootstrapping approach to realistically model annotator behavior. Leveraging datasets with a large number of ratings and persistent rater identifiers, we analyze the tradeoffs between the number of items ($N$) and the number of responses per item ($K$) required to achieve statistical significance.

Personalisation is a standard feature of conversational AI systems used by millions; yet, the efficacy of personalisation methods is often evaluated in academic research using simulated users rather than real people. This raises questions about how users and their simulated counterparts differ in interaction patterns and judgements, as well as whether personalisation is best achieved through context-based prompting or weight-based fine-tuning. Here, in a large-scale within-subject experiment, we re-recruit 530 participants from 52 countries two years after they gave their preferences in the PRISM dataset (Kirk et al., 2024) to evaluate personalised and non-personalised language models in blinded multi-turn conversations. We find preference fine-tuning (P-DPO, Li et al., 2024) significantly outperforms both a generic model and personalised prompting but adapting to individual preference data yields marginal gains over training on pooled preferences from a diverse population. Beyond length biases, fine-tuning amplifies sycophancy and relationship-seeking behaviours that people reward in short-term evaluations but which may introduce deleterious long-term consequences. Replicating this within-subject experiment with simulated users recovers aggregate model hierarchies but simulators perform far below human self-consistency baselines for individual judgements, discuss different topics, exhibit amplified position biases, and produce feedback dynamics that diverge from humans.

Efficient transfer learning methods for large-scale vision-language models ($e.g.$, CLIP) enable strong few-shot transfer, yet existing adaptation methods follow a fixed fine-tuning paradigm that implicitly assumes a uniform importance of the image and text branches, which has not been systematically studied in image classification. Through extensive analysis, we reveal a Branch Bias issue in vision-language image classification: adapting the image encoder does not always improve performance under out-of-distribution settings. Motivated by this observation, we propose A$_3$B$_2$, an Adaptive Asymmetric Adapter that alleviates Branch Bias in few-shot learning. A$_3$B$_2$ introduces Uncertainty-Aware Adapter Dampening (UAAD), which automatically suppresses image-branch adaptation when prediction uncertainty is high, enabling soft and data-driven control without manual intervention. Architecturally, A$_3$B$_2$ adopts a lightweight asymmetric design inspired by mixture-of-experts with Load Balancing Regularization. Extensive experiments on three few-shot image classification tasks across 11 datasets demonstrate that A$_3$B$_2$ consistently outperforms 11 competitive prompt- and adapter-based baselines.

Text-to-image (T2I) generative models are increasingly used to produce content for education, media, and public-facing communication, and are starting to be integrated into higher-impact pipelines. Since generated images tend to reinforce stereotypes, producing representational erasure via "default" depictions and shaping perceptions of who belongs in certain roles, a growing body of work has proposed metrics to quantify gender bias in T2I outputs. Yet existing evaluations remain fragmented. Metrics are often reported without a shared view of what they measure, what assumptions they entail, or how their results should be interpreted under different deployment contexts. This limits the usefulness of gender bias measurement for both technical auditing and emerging governance discussions. We propose a risk-aligned auditing framework for gender bias in T2I models composed of three constituents that connects risk categories, evaluation metrics, and harms. First, we identify risk-tiered use-case profiles aligned with the EU AI Act's risk categories to motivate why auditing expectations may vary with deployment contexts and stakeholder exposure. Second, we construct a metric catalog that consolidates gender-bias evaluation methods and organizes them in three measurement categories: gender prediction, embedding similarity, and downstream task. Third, we introduce a harm typology that maps context-dependent harm categories (e.g., representational, quality-of-service) to specific risk-tired scenarios. Finally, we introduce THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias) that help formulate auditing systematically by the incorporation of context, scenario and bias manifestation, harm hypotheses, and audit strategy.

Neural operators learn to map initial conditions to the terminal solution of partial differential equations (PDEs), providing a surrogate for the full operator mapping. This enables rapid prediction across different input configurations. While recent neural operator architectures have demonstrated strong performance on diverse PDE tasks, their behavior under structured distribution shifts remains insufficiently understood. To investigate this, we study operator learning in a wave propagation setting governed by a one-dimensional variable-coefficient wave equation, using two representative architectures, the Fourier Neural Operator (FNO) and the Deep Operator Network (DeepONet). To examine their generalization under distribution shifts, we consider structured out-of-distribution (OOD) settings that independently vary input frequency and coefficient smoothness. The results show that under smoothness shifts, both models maintain stable performance, with FNO achieving lower error. In contrast, under frequency shifts, FNO exhibits a sharp increase in error under unseen high-frequency inputs, whereas DeepONet shows milder degradation despite higher overall error. Our analysis reveals that these differences arise from how each architecture represents and responds to variations in frequency structure. Together, these findings highlight a fundamental gap between strong in-distribution performance and generalization under distribution shifts in operator learning, underscoring the role of architectural representation bias in developing more reliable neural operators for physics-based PDE simulations beyond the training distribution.

Large language models (LLMs) are increasingly used as substitutes for human subjects in behavioral simulations, including synthetic social network generation. Yet it remains unclear how their relational outputs depend on prompt design, cultural framing, prompt language, and model scale. Building on homophily theory and structural balance theory, we formalize four LLM-based tie-formation mechanisms: sequential, global, local, and iterative, and treat them as distinct conditional distributions over edge sets. Using a fixed roster of 50 demographically grounded personas, we generate 192 verified directed networks across four cultural contexts, four prompt languages, three GPT-4.1 variants, and four prompting architectures, with two seeds per condition. We find that cultural framing shifts inbreeding homophily and largest-component connectivity. Political affiliation dominates tie formation under three methods, while the global method substitutes age, showing that prompt architecture functions as a substantive sociological variable. Model scale produces a stable divergence ranking, with the smallest variant behaving qualitatively differently rather than merely noisily. Prompt language alone sharply shifts religion homophily, especially under Hindi prompting, while leaving political homophily nearly invariant. LLM-generated networks match real social graphs on clustering and modularity better than standard graph baselines, yet encode demographic biases above empirical levels. These results show that prompt choices often treated as implementation details encode substantive sociological assumptions.

Discrepancies between an agent's actual knowledge and what a person thinks the agent knows can hinder interactions. If an agent could detect such discrepancies, it could provide feedback to account for them and improve current and future interactions. Using the I-POMDP as a framework for a second-order Theory of Mind (ToM-2), this work endows an agent with the ability to model the evolution of a person's erroneous beliefs about an agent and the cognitive biases and heuristics (CBH) from which they arise. In doing so, the agent can detect when CBH might be at play during an interaction and adaptively generate feedback that accounts for them. An in-person user study shows how a ToM-2 learner can account for the effects of a teacher's CBH to significantly improve the informativeness of teacher actions, and subjective results suggest people find the ToM-2 learner's feedback more useful.

Large Language Models (LLMs) can generate fluent political text at scale, raising concerns about synthetic discourse during crises and social conflict. Existing AI-text detection often focuses on sentence-level cues such as perplexity, burstiness, or token irregularities, but these signals may weaken as generative systems improve. We instead adopt a Computational Social Science perspective and ask whether synthetic political discourse behaves like an observed online population. We construct a paired corpus of 1,789,406 posts across nine crisis events: COVID-19, the Jan. 6 Capitol attack, the 2020 and 2024 U.S. elections, Dobbs/Roe v. Wade, the 2020 BLM protests, U.S. midterms, the Utah shooting, and the U.S.-Iran war. For each event, we compare observed discourse from social platforms with synthetic discourse generated for the same context. We evaluate four dimensions: emotional intensity, structural regularity, lexical-ideological framing, and cross-event dependency, using mean gaps and dispersion evidence. Across events, synthetic discourse is fluent but population-level unrealistic. It is generally more negative and less dispersed in sentiment, structurally more regular, and lexically more abstract than observed discourse. Observed discourse instead shows broader emotional variation, longer-tailed structural distributions, and more context-specific, colloquial lexical markers. These differences are event-dependent: larger for fast-moving, decentralized crises and smaller for formal or institutionally mediated events. We summarize them with a simple event-level measure, the Caricature Gap. Our findings suggest that the main limitation of synthetic political discourse is not grammar or fluency, but reduced population realism. Population-level auditing complements traditional text-detection and provides a CSS framework for evaluating the social realism of generated discourse.

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.

Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.

Large Language Models (LLMs) suffer from order bias, where their performance is affected by the arrangement order of input elements. This unfairness limits the model's applications in scenarios such as in-context learning and Retrieval-Augmented Generation (RAG). Recent studies attempt to obtain optimal or suboptimal arrangements based on statistical results or using dataset-based search, but these methods increase inference overhead while leaving the model's inherent order bias unresolved. Other studies mitigate order sensitivity through supervised fine-tuning using augmented training sets with multiple order variants, but often at the cost of accuracy, trapping the model in consistent yet incorrect hallucinations. In this paper, we propose \textbf{D}ual \textbf{G}roup \textbf{A}dvantage \textbf{O}ptimization (\textbf{DGAO}), which aims to improve model accuracy and order stability simultaneously. DGAO calculates and balances intra-group relative accuracy advantage and inter-group relative stability advantage, rewarding the policy model for generating order-stable and correct outputs while penalizing order-sensitive or incorrect responses. This marks the first time reinforcement learning has been used to mitigate LLMs' order sensitivity. We also propose two new metrics, Consistency Rate and Overconfidence Rate, to reveal the pseudo-stability of previous methods and guide more comprehensive evaluation. Extensive experiments demonstrate that DGAO achieves superior order fairness while improving performance on RAG, mathematical reasoning, and classification tasks. Our code is available at: https://github.com/Hyalinesky/DGAO.

Modern Earth System Models (ESMs) operate on horizontal scales far larger than typical cloud features, requiring stochastic subcolumn generators to represent subgrid horizontal and vertical cloud variability. Traditional physically-based generators often rely on analytical cloud overlap paradigms, such as exponential-random decorrelation, which can struggle to capture the complex, anti-correlated behavior of non-contiguous cloud layers. In this study, we introduce a novel two-stage machine learning subcolumn generator for the GEOS atmospheric model, utilizing a Conditional Variational Autoencoder combined with a Generative Adversarial Network (CVAE-GAN) and a U-Net architecture. Trained on a merged CloudSat-CALIPSO height-resolved cloud optical depth dataset, the ML generator creates 56 stochastic subcolumns representing cloud occurrence and optical depth profiles. Evaluated against the established Räisänen, the ML approach accurately reproduces bimodal cloud overlap distributions, significantly reduces biases in grid-mean statistics, and halves the root-mean-square error in ISCCP-style cloud-top pressure and optical thickness joint histograms. The improvements brought by our deep generative models translate into more accurate offline radiative transfer calculations, reducing the global-mean shortwave top-of-atmosphere cloud radiative effect bias by a factor of three. Provided that the generator can be accelerated on CPUs, this offers a practical pathway to reduce structural errors at the cloud-radiation interface.

This paper examines what it means for a medical AI system to be right by grounding the question in a specific clinical context: the automatic classification of plasma cells in digitized bone marrow smears for the diagnosis of multiple myeloma. Drawing on philosophy of science and research ethics, the paper argues that correctness in medical AI is not a singular property reducible to benchmark performance, but a multi-dimensional concept involving the availability of expertly labeled medical datasets, the explainability and interpretability of model outputs, the clinical meaningfulness of evaluation metrics, and the distribution of accountability in human-AI workflows. As such, the paper develops this argument through four interrelated themes: the instability of ground truth labels, the opacity of overconfident AI, the inadequacy of standard clinical metrics, and the risk of automation bias in time-pressured clinical settings.

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.

Autoregressive language models are widely used for text evaluation, however, their left-to-right factorization introduces positional bias, i.e., early tokens are scored with only leftward context, conflating architectural asymmetry with true text quality. We propose masked reconstruction as an alternative paradigm, where every token is scored using full bidirectional context. We introduce DiffScore, an evaluation framework built on Masked Large Diffusion Language Models. By measuring text recoverability across continuous masking rates, DiffScore eliminates positional bias and naturally establishes an evaluation hierarchy from local fluency to global coherence. We further provide diagnostic tools unavailable to autoregressive frameworks: multi-timestep quality profiles that decompose scores across masking rates, and bidirectional PMI decomposition that disentangles fluency from faithfulness. Experiments across ten benchmarks show that DiffScore consistently outperforms autoregressive baselines in both zero-shot and fine-tuned settings. The code is released at: https://github.com/wenlai-lavine/DiffScore.

Multimodal Large Language Models (MLLMs) have shown strong performance in multi-image cross-modal retrieval, yet suffer from severe position bias, where predictions are dominated by input order rather than semantic relevance. Through empirical analysis, we identify a phenomenon termed Logit-Attention Divergence, in which output logits are heavily biased while internal attention maps remain well-aligned with relevant visual evidence. This observation reveals a fundamental limitation of existing logit-level calibration methods such as PriDe. Based on this insight, we propose a training-free, attention-guided debiasing framework that leverages intrinsic attention signals for instance-level correction at inference time, requiring only a minimal calibration set with negligible computational overhead. Experiments on MS-COCO-based benchmarks show that our method substantially improves permutation invariance and achieves state-of-the-art performance, enhancing accuracy by over 40% compared to baselines. Code is available at https://github.com/brightXian/LAD.

Average-reward reinforcement learning requires estimating the gain and the bias, which is defined only up to an additive constant. This makes direct distributional analogues ill-posed on the real line. We introduce a quotient-space formulation in which state-indexed bias laws are identified up to a common translation, together with a categorical parameterization that respects this symmetry. On this quotient-categorical space, we define a projected average-reward distributional operator and show that it is well-defined, non-expansive in a coordinate Cramér metric, and admits fixed points. We then study sampled recursions whose mean-field maps are asynchronous relaxations of this operator. In an idealized centered-reward setting, a one-state temporal-difference update enjoys almost sure convergence together with finite-iteration residual bounds under both i.i.d. and Markovian sampling. When the gain is unknown, we augment the recursion with an online gain estimator, and prove non-expansiveness and Markovian convergence of the resulting coupled scheme. Finally, we show that synchronous exact updates are gain-independent at the quotient-law level, isolating a structural contrast between ideal quotient distributions and practical fixed-grid categorical representations.

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.

Preference learning methods like Direct Preference Optimization (DPO) are known to induce reliance on spurious correlations, leading to sycophancy and length bias in today's language models and potentially severe goal misgeneralization in future systems. In this work, we provide a unified theoretical analysis of this phenomenon, characterizing the mechanisms of spurious learning, its consequences on deployment, and a provable mitigation strategy. Focusing on log-linear policies, we show that standard preference-learning objectives induce reliance on spurious features at the population level through two channels: mean spurious bias and causal-spurious correlation leakage. We then show that this reliance creates an irreducible vulnerability to distribution shift: more data from the same training distribution fails to reduce the model's dependence on spurious features. To address this, we propose tie training, a data augmentation strategy using ties (equal-utility preference pairs) to introduce data-driven regularization. We demonstrate that this approach selectively reduces spurious learning without degrading causal learning. Finally, we validate our theory on log-linear models and provide empirical evidence that both the spurious learning mechanisms and the benefits of tie training persist for neural networks and large language models.

Large language models increasingly mediate decisions that turn on moral judgement, yet a growing body of evidence shows that their implicit preferences are not culturally neutral. Existing cultural alignment methods either require per-country preference data and fine-tuning budgets or assume white-box access to model internals that commercial APIs do not expose. In this work, we focus on this realistic black-box, public-data-only regime and observe that within-country sociodemographic disagreement, not consensus, is the primary steering signal. We introduce DISCA (Disagreement-Informed Steering for Cultural Alignment), an inference-time method that instantiates each country as a panel of World-Values-Survey-grounded persona agents and converts their disagreement into a bounded, loss-averse logit correction. Across 20 countries and 7 open-weight backbones (2B--70B), DISCA reduces cultural misalignment on MultiTP by 10--24% on the six backbones >=3.8B, and 2--7% on open-ended scenarios, without changing any weights. Our results suggest that inference-time calibration is a scalable alternative to fine-tuning for serving the long tail of global moral preferences.

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.

Muon has emerged as an efficient alternative to Adam for pretraining, yet remains underused for fine-tuning. A key obstacle is that most open models are pretrained with Adam, and naively switching to Muon for fine-tuning leads to degraded performance due to an optimizer mismatch. We investigate this mismatch through controlled experiments and relate it to the distinct implicit biases of Adam and Muon. We provide evidence that the mismatch disrupts pretrained knowledge, and that this disruption scales with update strength. This leads us to hypothesize that constraining updates should mitigate the mismatch. We validate this with LoRA: across language and vision tasks, LoRA reduces the performance gap between Adam and Muon observed under full fine-tuning. Studies on LoRA rank, catastrophic forgetting, and LoRA variants further confirm that mismatch severity correlates with update strength. These results shed light on how optimizer mismatch affects fine-tuning and how it can be mitigated. Our code is available at https://github.com/XingyuQu/muon-finetune.

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.

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9% to 81.1% at 16K; and on GPQA with agentic tool use, it yields a 24% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529$\times$ and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT

Machine learning algorithms are being used in high-stakes decisions, including those in criminal justice, healthcare, credit, and employment. The research community has responded with two largely independent research fields: \emph{algorithmic fairness}, which targets equitable outcomes, and \emph{explainable AI} (XAI), which targets interpretable reasoning. This survey identifies and maps a novel blind spot at their intersection, which is a model that can satisfy every standard fairness criterion in its outputs while being profoundly unfair in its \emph{reasoning process}. We refer to this as the procedural bias, and mitigating it requires treating the fairness of explanations as a distinct object of scientific study. To our knowledge, we provide the first unified theoretical and literature review of this emerging field and elucidate the drawbacks of post-hoc explainers in certifying explanation fairness. Our central contribution is a \emph{conditional invariance framework} formalizing explanation fairness as the requirement that explanations should be indifferent regardless of the protected attributes $P(E(X) \in \cdot \mid X_\text{rel} = x_\text{rel},\, A = a) = P(E(X) \in \cdot \mid X_\text{rel} = x_\text{rel},\, A = b)$ for all task-relevant $x$, a single principle from which all existing explanation fairness metrics emerge as partial operationalizations. We introduce a seven-dimensional taxonomy, identify three generative mechanisms of explanation inequity (representation-driven, explanation-model mismatch, actionability-driven), and propose a canonical six-step evaluation workflow for operationalizing explanation fairness audits in practice.

We use sparse autoencoder (SAE) feature steering to amplify Dark Triad personality traits (Machiavellianism, narcissism, and psychopathy) in Llama-3.3-70B-Instruct and evaluate the resulting behavioral changes across five psychological instruments. The steered model becomes substantially more exploitative, aggressive, and callous on novel behavioral scenarios (d=10.62) while its cognitive empathy remains intact, reproducing the empathy dissociation characteristic of human Dark Triad populations. Critically, strategic deception is completely unaffected across all features, suggesting that exploitation and deception may operate through dissociable computational pathways in large language models. Individual feature analysis reveals non-redundant encoding, with each feature driving distinct antisocial mechanisms through separable computational pathways. We also show that feature discovery method itself modulates intervention depth: contrastively-discovered features change both self-report and behavior, while semantically-searched features change only self-report (d=12.65 between methods on behavior). These findings suggest that antisocial tendencies in at least one large language model comprise dissociable components rather than a unified construct, with implications for how such tendencies should be detected, measured, and controlled.

Discrete diffusion models generate structured sequences by progressively unmasking tokens, but enforcing global property constraints during generation remains an open challenge. We propose primal-dual guided decoding, an inference-time method that formulates constrained generation as a KL-regularised optimisation problem and solves it online via adaptive Lagrangian multipliers. At each denoising step, the method modifies token logits through an additive, constraint-dependent bias, with multipliers updated by mirror descent based on constraint violation. The bias arises as the optimal KL-regularised projection of the constraint, so the constrained distribution remains as close as possible to the model's unconstrained distribution while still satisfying the constraint. The method requires no retraining and no additional model evaluations beyond standard sampling, supports multiple simultaneous constraints, and provides formal bounds on constraint violation. We evaluate our approach on topical text generation, molecular design, and music playlist generation, showing that a single algorithm instantiated via domain-specific scoring functions improves constraint satisfaction while preserving relevant domain-specific quality metrics.