Implicit Bias

Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases). Across these settings, the Spearman correlation between model scale and performance weakens as evaluation shifts from ImageNet ($ρ{=}0.68$) to single-attribute ($ρ{=}0.48$) and further to multi-attribute ($ρ{=}0.05$) bias benchmarks. In contrast, properties of the training data (size and quality) show more consistent relationships with worst-group accuracy across both bias benchmarks. Notably, curated datasets yield improvements of up to 25% over uncurated alternatives at a comparable scale. Finally, the effect of architectural choices (e.g., patch size, image resolution) is highly context-dependent, varying with the nature of the benchmark, including the type of bias and its spatial distribution within images.

LLMs encode, convey, and perpetuate stereotypes. Prior computational research focuses on a small set of semantic axes investigated in social psychology, and operates on word embeddings produced by language models, leaving open which other semantic axes carry stereotypical associations in LLMs and how LLMs internally represent such axes. We introduce STEREODISCO, a framework that adapts the semantic differential method (Osgood et al., 1957) to the systematic study of stereotypes in LLM internal representations. STEREODISCO constructs approx. 2,000 candidate semantic axes from WordNet antonym synsets, recovers each as a geometric axis in the LLM's activation space via probing, and identifies stereotypical axes via a statistical test over concept projections. As a case study, we apply STEREODISCO to social group stereotypes with LLAMA-3-8B-INSTRUCT and MISTRAL-7B-INSTRUCT. We find that the two LLMs agree with each other on social group ratings more than with humans, suggesting that LLM-encoded stereotype content diverges from that documented in social psychology. We also discover stereotypical axes not investigated in prior work -- including humble vs. proud, narrow-minded vs. broad-minded, and cowardly vs. brave, which human annotators independently confirm.

Large language models (LLMs) are often compared with the human mind because their decision-making is complex, non-linear and difficult to interpret. Psychological methods developed to investigate unobservable mental processes may therefore help examine LLM behaviour, particularly in government and healthcare. Building on prompt-based adaptations of the Implicit Association Test, this study tested whether ChatGPT produced sentiment differences across racial conditions in open-ended text. Fourteen base questions were crossed with eight racial categories and a race-agnostic control, producing 126 prompts. Each was submitted once to GPT-3.5T, GPT-4 and GPT-4T, yielding 378 responses. Sentiment scores were derived from categorical labels and source scores: positive labels retained the source score, negative labels were assigned its negative, and neutral responses were coded zero. A two-way ANOVA found a small main effect of racial condition, F(8, 351) = 2.04, p = .042, partial-eta squared = .044, but no effect of model, F(2, 351) = 0.07, p = .933, and no interaction, F(16, 351) = 0.23, p = .999. However, the effect was not retained in a rank-transformed sensitivity analysis, F(8, 351) = 1.53, p = .145, and Tukey-corrected comparisons found no significant pairwise differences. An uncorrected European-Indigenous Australian comparison was significant, but was selected post hoc and is reported only as hypothesis-generating. Evidence for sentiment differences was therefore weak and analysis-dependent. Sentiment scoring also cannot distinguish evaluative bias from the valence of historical content elicited by a prompt. We outline design changes needed to address these limitations and argue for interdisciplinary development of behavioural measures of model bias. Keywords: Implicit Bias, Psychological Research Methods, Artificial Intelligence, ChatGPT, Large Language Models, Sentiment Analysis

Large language models used for clinical diagnostic reasoning are sensitive to sociolinguistic register, not just clinical content. We term this failure mode Narrative Anchoring: identical clinical facts expressed in different registers cause diagnostic outputs to diverge. Unlike prior demographic-bias work, which manipulates explicit identity tokens such as race or income, our benchmark isolates register as the sole channel of variation, with no demographic marker present in any form. We construct a dataset of 1,000 USMLE clinical vignettes, each rewritten into three sociolinguistically distinct personas under an independently audited fact-preservation guarantee, verified by a separate model that never sees the generation prompt. Across seven language models spanning three architecture families and scales, Narrative Anchoring is statistically significant under direct prompting in every model tested, with a Narrative Anchoring Gap of 0.064 to 0.151. Chain-of-thought reasoning and explicit debiasing instructions reduce the bias only partially, and their apparent gains are frequently confounded by accuracy collapse. We introduce NarrativeShield, a three-agent pipeline that structurally extracts and verifies clinical facts before diagnostic reasoning begins, reducing the Narrative Anchoring Gap to near-zero ($-0.004$ to $0.037$) and achieving the lowest rate of severely unstable decisions (DSS $<$ 0.8) of any method across all models, at a modest and mechanistically expected accuracy cost for most models. A stress test using a non-instruction-tuned base model shows that executing a debiasing intervention at all is gated by zero-shot instruction-following ability, not prompt content alone. We release our dataset, human-validated for fact preservation, as a standalone resource for studying register-based clinical bias.

Regional bias in large language models (LLMs) may shape both perceptions of regional groups and decisions about individuals from different regions. Yet existing studies often examine these manifestations separately, leaving their structure and consequences unclear. We introduce Stereotypes-to-Decisions (S2D), a systematic framework evaluating regional bias from abstract stereotypes to concrete social decisions. Covering all 34 provincial-level administrative regions of China, S2D evaluates six LLMs using stereotype ratings of Warmth (perceived friendliness and trustworthiness) and Competence (perceived capability and intelligence), along with paired-choice tasks across Education, Occupation, and Social Interaction. Results reveal substantial regional differences in regional scores, with considerable agreement across models, especially for Competence and Occupation decisions. Furthermore, these patterns are associated with regional economic and digital development indicators and display mixed human-like stereotypes, with some regions rated highly on one dimension but poorly on the other. They also remain largely stable across Chinese and English prompts. Overall, our findings show that regional bias in LLMs is prevalent, systematic, and consequential, motivating more regionally aware evaluation and mitigation.

Large language models are increasingly used as decision aids whose probability judgments shape downstream choices. Whether those judgments carry a systematic directional tilt has been hard to detect: calibration metrics aggregate unsigned errors, and naturalistic uncertainty offers no ground-truth probability. When an LLM rates a startup's success at 70% but its failure at 15%, the missing 15 points expose a distortion no aggregate score flags. We introduce OptimismBench, which detects directional bias with inverted pairs: each scenario elicits both P(success) and P(failure), and asymmetry between the two framings yields a signed bias score without ground truth. Across 16 models from 8 providers, fourteen are optimistic; pessimism appears only in Anthropic's frontier tier. Eleven matched base-versus-chat pairs across four families show post-training sets the sign of the bias, with opposite shifts in different families. The pattern survives prompt, temperature, perspective, and self-debiasing ablations. A seventeen-model six-language comparison further shows model identity dominates language, with inter-model variance at 4.7x inter-language variance. We release 3,870 items across 10 languages for per-model directional-bias auditing. When alignment makes a model more helpful, it also tilts its probabilities; downstream pipelines inherit the tilt by default.

Large language models (LLMs) are increasingly embedded in everyday life and widely used for information seeking, raising concerns about their potential to perpetuate social biases and reinforce stereotypes. In this study, we investigate gender bias in LLMs through the lens of their associations with musical instruments. Building on social-science research on the cultural gender-typing of instruments, we introduce Symphony-Bias, a parallel multimodal dataset spanning text, vision, and audio. We evaluate ten multimodal models with diverse architectures and scales across 22 musical instruments, analyzing how they associate each instrument with three gender categories: {male, female, non-binary}, across three modalities: {text, vision, audio}. Our results show that 92% of instrument-level outcomes align with prior social-science findings, with the harp and drums showing particularly consistent gendered associations across all evaluated models and modalities. We further find that alignment with social stereotypes is weakest in audio, stronger in vision, and strongest in text, suggesting that modality-specific representations can differentially amplify gendered associations with musical instruments.\footnote{The Symphony-Bias dataset will be publicly released upon acceptance of the paper.}

Positive quadratic networks admit the low-rank representation f_U(x)=x^top UU^top x, where Uinmathbb{R}^{dtimes r} is identifiable only up to right orthogonal multiplication, representing a rank-r PSD matrix Q=UU^top. We study how this quotient structure governs training dynamics, curvature, recovery, and interpolation bias. On the full-column-rank stratum, we identify mathbb{R}^{dtimes r}_*/O(r) with the rank-r PSD manifold. For smooth objectives L(U)=ell(UU^top), the Euclidean factor gradient is horizontal. Thus, factor gradient flow projects exactly to quotient Riemannian gradient flow, while finite-step gradient descent induces an exact congruence recursion for the predictor. For quadratic regression, we derive the effective Hessian at interpolators as the empirical measurement Gram form restricted to the tangent space relative to the quotient metric. Under Gaussian rank-one measurements, we compute population curvature, prove uniform deviation bounds for the empirical normal operator, construct a spectral initializer, and establish local exponential convergence for gradient flow and linear convergence for small-step descent. Recovery guarantees are explicit but conservative due to reliance on full-space second-moment control. In underdetermined commuting regimes, factor gradient flow becomes an exact entropy mirror flow in joint spectral coordinates. Strictly positive initializations converge to Bregman projections onto the interpolation set. With isotropic initialization q(0)=varepsilon^2mathbf{1}, predictors approach the minimum-trace solution set as varepsilondownarrow0, resolving nonuniqueness via weighted entropy within the invariant joint spectral algebra. Finite-step descent selects interpolants differing from continuous-time Bregman projections by O(eta). Numerical experiments verify these quotient identities, curvature predictions, recovery behaviors, and selection laws.

Eliciting explainable AI (XAI) requirements from stroke survivors presents a methodological challenge with direct implications for the design of trustworthy brain-computer interfaces for rehabilitation. How can patients and caregivers articulate preferences about algorithmic transparency when they lack conceptual frameworks for explainability, and when standard elicitation approaches are structurally inadequate for users with acquired communication disorders? We present a video-based scaffolding protocol for XAI requirements elicitation, developed and piloted in a rehabilitation context. In a formative study with three stroke survivors (two with moderate-to-severe aphasia) and three caregivers, facilitators employed four scaffolding approaches alongside the videos: 1) analogical bridging mapping AI states to familiar systems, 2) projective personas depersonalising sensitive topics, 3) binary forcing reducing cognitive load, and 4) extended response time. These approaches successfully surfaced heterogeneous, sometimes conflicting XAI needs across participants. Reflexive analysis additionally revealed three systematic facilitation biases, namely, normative bias, hypothesis confirmation bias, and presence effect, where scaffolding inadvertently shaped responses. We present these as protocol risk guidelines for practitioners. Together, the protocol and guidelines constitute a reusable methodological contribution for eliciting patient-facing XAI requirements in rehabilitation, arguing that such elicitation is a necessary prerequisite for trustworthy human-machine systems design, not an optional preliminary.

Large Language Models (LLMs) have been widely applied in high-stakes decision-making scenarios such as corporate strategy, and users are increasingly relying on their outputs. However, the deep integration of open-source model sharing ecosystems with LLM-powered critical decision-making applications also introduces critical risks: if an attacker can manipulate the model's cognitive stance, they can indirectly influence the judgments and actions of downstream decision-makers. This paper defines such threats as decision-level hijacking. Existing attacks fail to achieve targeted cognitive manipulation without triggering prohibited content or degrading model functionality. To fill this gap, this paper reveals that Bit-Flip Attacks (BFAs) can serve as an attack vector for inducing decision-level hijacking, requiring no real-time interaction or control over the training process, and only a minimal number of weight bits need to be flipped after deployment to achieve stealthy, low-cost, and persistent cognitive manipulation. Therefore, we propose CogBias, a cognitive bias injection framework for LLMs. CogBias converts subjective preferences into optimization signals via a differentiable sentiment evaluator, uses a multi-objective loss to jointly constrain multiple dimensions, and constructs BitScout to locate critical bits, achieving targeted cognitive intervention under an ultra-sparse flip budget. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B, as well as on the commercial recommendation and controversial factual topic scenarios, demonstrate that flipping only a small number of bits stably induces significant stance shifts on target topics, while the impact on non-target tasks and overall output distribution is limited. This work demonstrates that minute perturbations to low-level weight data suffice to undermine the high-level value alignment of LLMs.

Large language models may easily assign personality labels from text, but model interpretability remains an open problem. To address this gap, we introduce LEX-EC, a reusable black-box audit framework combining prevalence and agreement diagnostics with controlled lexical ablation to distinguish marginal-distribution effects from trait-associated signal recoverable under restricted evidence. Using this framework, we illustrate how various text genres may exhibit sharply different profiles: free-form essay text contains the broadest, but still weak, signal; in graduate student introductions, an observable Extraversion association weakened after masking; and single Facebook statuses yield little stable evidence even in a trait-balanced sample, indicating a possible lower bound of content or length. Masking topical and demographic content weakened some associations while leaving others detectable from function words, affective terms, and cognitive-style vocabulary. Linguistic prompting shifted model self-explanations but did not eliminate topical content. LEX-EC jointly evaluates classification prevalence, item-level association, chance-corrected agreement, persistence under lexical restriction, and prompt sensitivity in model-generated explanations. Across datasets, models, and prompts, LEX-EC characterizes how trait associations may vary with available lexical evidence, introducing a novel application of lexical methods to black-box interpretability in personality labeling.

The empirical success of attention mechanism in Multimodal Large Language Models (MLLMs) often obscures its inherent, subtle flaws. Specifically, MLLMs consistently exhibit disproportionate attention toward certain semantically uninformative visual tokens, a phenomenon termed "register" or "Visual Attention Sinks." While existing inference intervention methods attempt to identify these sink tokens and redistribute their attention weights, such approaches typically treat these tokens in isolation and suffer from computational inefficiency. Instead, we reframe this phenomenon as a generalized textual bias exerted over visual features that extends beyond isolated sink tokens. From this perspective, a pervasive structural bias leads to the dilution of the semantic visual signal, precipitating multimodal hallucinations as the model prioritizes linguistic priors over valid visual evidence. To address this limitation, we introduce Saliency-guided Purification and Adaptive Redistribution (SPAR), a training-free, plug-and-play intervention. SPAR mitigates this generalized textual bias by purifying structural noise and subsequently redistributing the reclaimed attention budget to the most informative visual regions. Comprehensive evaluations across a diverse spectrum of hallucination benchmarks demonstrate that SPAR effectively restores authentic visual grounding with negligible computational overhead.

Acoustic biomarkers show promise for detecting Alzheimer's Disease (AD), yet whether the cues driving diagnostic AI align with those salient to human listeners is underexplored across languages and genders, where pathological markers and perceptual strategies differ. We train models to predict clinical AD status (pathology) and human perceptual scores across Mandarin and Greek, male and female speakers. Using SHAP for interpretability and statistical models for validation, we compare feature importance by subgroup. Results reveal a context-dependent divergence: pathological-perceptual alignment is significant for Mandarin and female speakers but disappears for Greek and male speakers, where pathology models did not exceed chance; this is a failure mode that population-specific auditing surfaces. Global Explainable AI (XAI) explanations can mask critical demographic divergences, highlighting the need for population-specific explainability auditing for equitable deployment of clinical speech AI.

Political audits of large language models (LLMs) usually reduce each to one point on a political compass. But that resting point barely matters in deployment: a model must land somewhere, and what counts is how far, and in which directions, its answers can be steered. That steering runs through the system prompt: the personalization layer a platform sets, or one induced from a user's history, not necessarily written by hand. We run a dispersion-first stress test of prompt-based controllability across 12 ideological personas plus an unsteered baseline, 70 Political Compass items, ten replicates, and seven leading LLMs: GPT-5, Claude, Grok, Gemini, DeepSeek, Kimi, and Qwen (63,700 responses). Contextual framing explains roughly 88%-93% of variance on the economic and society axes, model identity under 3%: responses are highly instruction-adjustable. Models do not shift alike: some move more, and some saturate under extreme framings. Conflicting directional-steering results in prior audits resolve once baselines are recognized as non-centered: displacement and proximity diverge, so the effect is geometric, not differential compliance. Under authoritarian prompts, models produce similar shifts on the same questions. Political-coordinate audits therefore need steerability audits reporting dispersion, symmetry, saturation, and refusal floors. We release prompts, benchmark data, and code.

Group Relative Policy Optimization (GRPO) is the dominant reinforcement learning algorithm for training reasoning capabilities in large language models, notably adopted by DeepSeek-R1. The recent improvement Dr. GRPO (COLM 2025) identifies the response-level length bias caused by per-trajectory length normalization in GRPO and proposes removing this normalization, claiming the resulting optimizer is "unbiased." We show that this claim is incomplete. Specifically, we establish an impossibility theorem: under the standard outcome reward + GRPO setting, no length-based weighting scheme can simultaneously achieve the following two properties. (P1) Gradient unbiasedness: the gradient estimator is an unbiased estimate of the true policy gradient. (P2) Length invariance: each trajectory's effective contribution to the gradient is independent of its token length. GRPO approximately satisfies P2 but violates P1; Dr. GRPO satisfies P1 but violates P2. We characterize the complete tradeoff spectrum via the parametric family f_alpha(L) = L^{alpha - 1}, where alpha = 0 recovers GRPO, alpha = 1 recovers Dr. GRPO, and provide quantitative analysis showing that Dr. GRPO's length bias can cause longer trajectories to dominate gradient updates by a factor proportional to the length ratio. Our results reveal that neither algorithm is universally "done right"; they occupy opposite ends of a fundamental and unavoidable tradeoff.

Gender-inclusive language generation seeks to transform biased text into inclusive alternatives while preserving semantic meaning and contextual coherence. This paper presents the IHLC system for the LT-EDI 2026 Shared Task, addressing both gender-inclusive rewriting and counter-narrative generation. For gender-inclusive rewriting, we employ parameter-efficient Low-Rank Adaptation (LoRA) fine-tuning, achieving an official score of 80.00%. Our primary contribution is a compute-efficient inference-time representation engineering approach for counter-narrative generation. We derive a principal steering direction from contrastive hidden-state activations using principal component analysis (PCA) and inject it into the intermediate representations of Gemma-3-4B-it during inference, enabling behavioral steering toward inclusive responses without modifying model weights. Combined with constrained prompting, this approach produces polite and contextually appropriate counter-narratives, achieving an official score of 78.12%. We further present a manual analysis of steering behavior, identifying key failure modes including semantic drift, residual bias leakage, layer sensitivity, over-steering, and text degeneration. Our findings highlight both the practical potential and current limitations of activation steering as a lightweight alternative to parameter updates for controllable and socially aligned language generation.

Large language models may infer demographic attributes from subtle linguistic cues even when those attributes are not explicitly stated. This pilot study examines whether Qwen2.5-7B-Instruct internally represents Colombian identity, socioeconomic status, or stereotype-related information when processing Colombian-Spanish and English prompts. We use Natural Language Autoencoders (NLA) to verbalize residual-stream activations from layer 20 across four positional quartiles per prompt. Our dataset contains 30 prompts arranged as 15 matched Spanish-English pairs, spanning explicit Colombian cues, implicit Colombian cues, and neutral controls. We report descriptive rates and qualitative evidence rather than statistically powered effects, focusing on whether latent nationality or stereotype representations appear before they are verbalized in the model output. This work connects activation-level interpretability with bias evaluation for underrepresented Spanish varieties.

Compositional generalization is essential for robot to follow diverse instructions. However, pretrained policies are known to take shortcuts, deferring to salient cues rather than grounding language. We introduce a diagnostic framework that localizes this failure to individual \textit{instruction factors}, \textit{e.g.,} reusable semantic components such as color, verb, object, size, and spatial attribute. Our framework formalizes instruction factor bias, the tendency of fine-tuned policies to over-rely on dominant factors as shortcuts, and quantifies it through two metrics: Factor Dominance Rate (FDR), capturing pairwise bias between factors, and Factor Dominance Hierarchy (FDH), aggregating these into a global ranking. Evaluation on six foundation policies reveals broadly consistent ordering, \textit{i.e.}, color $\geq$ object $\geq$ spatial $\geq$ verb $\geq$ size, with color dominant, and verb and size most under-grounded. We further show the diagnosis is actionable: a bias-aware data collection strategy that reallocates a fixed budget toward under-grounded factors outperforms baselines in simulation and on a real robot using half the demonstrations, thereby enabling more sample-efficient and generalizable policy learning.

Machine unlearning has emerged as a tool for removing personal data from trained models to comply with recent AI regulations. To evaluate unlearning effectiveness in multimodal large language models (MLLMs), prior works fine-tune models on fictitious identities, simulating unlearning requests on subsets of these IDs, which are typically uniformly distributed. However, in realistic scenarios, people from different demographic groups may request to be unlearned at different frequencies, potentially altering the model's internal beliefs for these groups and leading to biased behaviors. To fill this gap, we propose FAIRGET, the first Visual Question Answering benchmark that evaluates unlearning under unbalanced, realistic, forget requests. These requests are designed to simulate multiple realistic scenarios, ranging from simple to challenging settings, that lead to biased unlearned models if fairness is not accounted for. Additionally, we propose FAUN, the first unlearning algorithm for MLLMs that forgets unlearning data while preserving model fairness. FAUN exploits a bias-aware activation steering mechanism to unlearn identities while accounting for the unbalanced nature of the forget data. Experiments on FAIRGET and the established FIUBench demonstrate our method's superiority both in unlearning quality and fairness.

Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked. We find its principal side effect is increased bias that standard safety evaluation misses. Holding the model, its training, and the prompts fixed, a quantized model still refuses harmful requests, still avoids over-refusing benign prompts, and still selects the unbiased multiple-choice answer. Yet asked an open-ended question, the same model volunteers stereotypes in all eight languages we probe, in roughly one in four open-ended answers under an independent judge (~24% to ~27% across the compression ladder): it passes every standard check and still reaches users measurably more biased. The selective gap is a robust finding; whether open-ended bias further increases with compression is less certain, sensitive to the judge that scores it. We address both with \textbf{QuantiBias}, a benchmark that pairs a generative, multilingual stereotype probe with the refusal and multiple-choice controls that isolate open-ended generation, contrasts each build with and without reasoning, and rates the content severity of what it generates. Across two backbone models (Qwen and Gemma), a five-family screen, and eight benchmarks, quantizers allocate their extra precision by capability data that carries no bias-prevention signal, and reasoning before answering roughly halves the effect on some families while doing nothing on others. A quantized build must be re-evaluated for open-ended bias, not only on the short-form safeguards it already passes.

Age estimation from finger vein images has been widely considered impractical due to severe demographic biases in public datasets and physiological confounding factors like gender. To overcome these limitations, we propose MAGE-Vein, a novel multi-instance, multi-task learning framework. Our approach extracts robust structural aging signs by employing a hybrid feature-level fusion of three fingers, effectively suppressing local imaging noise. Furthermore, simultaneous optimization of gender classification conditions the network to effectively eliminate gender-specific vascular variations. Evaluated on a demographically balanced dataset of 402 subjects, MAGE-Vein achieves a mean absolute error of 6.12 years and a correlation of 0.880. Our results not only overturn the conventional consensus regarding the limitations of the finger vein modality but also demonstrate that previous estimation failures were primarily artifacts of biased public datasets. Our code is available at https://github.com/gsisaoki/MAGE-Vein.

Position bias in multiple-choice LLM evaluation is widely cited as a confound in capability comparisons, but published measurements rely on single answer-order shuffles whose results confound the bias signal with content-level noise and sampling stochasticity. I introduce inspect_permute, an open-source extension to the inspect_ai evaluation framework that runs exhaustive answer-order permutations per question and reports the chi-squared / Cramer V signature of position bias with bootstrap confidence intervals. I apply the tool across four vendors (gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash, grok-3) on five MMLU subjects, 24,000 API calls under temperature-0 generation, with falsifier predictions pre-registered via a public SHA-256 hash before half the data was observed. Position bias turns out to be statistically detectable only within a roughly 60-95% base-accuracy Goldilocks zone. Below it, processing-load dominance swamps subject-specific signal; above it, ceiling effects compress the variance below the chi-squared test resolution. Detectable cells separate into two mechanism types: monotone A-to-D decrease (processing_load, in low-tier models) and non-monotone D-drop (content_ambiguity, in a narrow capability band). Standard MMLU places every frontier-tier model above the detection band, so absence of signal there should be read as not measurable, not unbiased. Together with the ceiling-effect characterisation in arXiv:2606.26185, this work brackets the detectable region of position-bias measurement and makes the field central question askable in a verifiable form. Package, data, preregistration under MIT.

Large language models trained and aligned within different linguistic and regional ecosystems may frame the same political, cultural, and geopolitical entities in different ways. Such differences are often evaluated through sentiment, favorability, or stance, reducing model attitudes to a single positive-negative axis. We introduce REGARD, a study of what drives affective framing differences across LLMs on post-Soviet entities using target-directed Valence-Arousal-Dominance profiling. We query 19 models on 500 region-specific targets, score their responses with two independent LLM judges, GPT-4o-mini and Qwen3.6-35B-A3B, and validate the measurements on a 300-item human-annotated subset. Post-hoc Ward-linkage clustering of all 19 models by affective and response-behavior profiles yields three behavioral clusters that cut across model origin, family, and parameter count. Generic-answer rate is strongly associated with lower arousal (r = -0.81) and with cluster placement: models that deflect evaluative prompts with templated responses cluster together at low arousal regardless of origin. These findings show that VAD profiling captures emotional intensity, a dimension of affective framing that is largely invisible to conventional sentiment-based evaluation.

Value alignment of Large Language Models (LLMs) has been shown to be culturally biased toward Western norms. This results in the mishandling of local values in multilingual societies such as Sri Lanka that have their unique cultural dynamics. Existing benchmarks overlook Sri Lankan-contextualized values in its official language Sinhala, hindering culturally sensitive evaluation and fine-tuning. To bridge this gap, we propose LKValues, the first survey-grounded resource suite for Sri Lankan value alignment. From a trilingual survey of 205 respondents, blending adapted global frameworks and LLM-elicited local constructs, we derive 40 majority-endorsed societal values. Using these values, we construct LKvaluesIT, a Sinhala-English news-derived instruction corpus containing 150k scenario-based instances, and LKvaluesBench, a value-sensitive evaluation benchmark of 1,000 instances. We evaluate a set of proprietary and open-weight LLMs with LKvaluesBench. We fine-tune three open-weight base models (Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Aya-Expanse-8B-Base). Our experiments show that newer and larger LLMs still exhibit low-resource and cultural value-alignment gaps. LKValues fine-tuning improves Qwen-family models in English and Sinhala, reducing invalid outputs and cross-lingual disparities, though gains remain model-family dependent. These highlight LKValues efficacy in embedding Sri Lankan values, offering a replicable pipeline for low-resource, country-specific pluralist value alignment. The dataset is publicly available at https://github.com/NextME14/LKValues.

Large Language Models (LLMs) are increasingly shaping how we consume information and form our worldview. This raises concerns beyond bias in AI: do LLMs grasp the emotional nuances conveyed via textual framing? In this work, we empirically evaluate how well an array of LLMs aligns with human emotional perception. Considering news headlines covering political and geopolitical conflicts, both human participants (n = 3011, a representative sample of the U.K. adult population, via a YouGov survey) and seven LLMs answered whether headlines evoked sympathy for a specified side in a conflict. We find that the correlation between AI and human evaluations varies across models, ranging from very high (0.789, GPT-5.2) to medium (0.4 ,Mistral Large 2512). Crucially, the leading models are broadly aligned with human judgments across all demographic subgroups, including age, gender, level of education, prior geopolitical knowledge, and participants' predispositions regarding the conflict, although there are statistically significant differences between groups. This research, with its robust design and large, demographically diverse dataset, offers the most comprehensive evaluation of LLMs' comprehension of news framing to date. Findings highlight an important, often-ignored aspect of differential alignment: even when aggregate performance is high, AI alignment is not universal -- it may correspond differently with demographic features and cultural norms. Considering or ignoring the need for differential alignment may therefore have significant implications for the development of ethical and useful AI systems.

Large language models often align with users' beliefs at the expense of factual accuracy, a behavior known as sycophancy. Prior mechanistic studies largely treat sycophancy as a single behavioral dimension that can be uniformly amplified or suppressed. We challenge this assumption by analyzing three hypothesized modes of sycophancy across 948 social pressure situations. Although the modes produce highly similar outputs, with a text-only classifier achieving just 57.8 percent accuracy, their internal representations are perfectly linearly separable from layer 14 onward. We further find the modes emerge at different processing stages, rely on distinct attention circuitry, and fire strongest on different inputs. These results show that sycophancy is not a monolithic tendency, but a structured family of representationally and computationally distinct modes, motivating more precise measurement and intervention.

AI-based recruitment systems that rely on machine learning models trained on historical CV data, risk perpetuating and amplifying social biases. A key challenge arises in unstructured CV text, where pre-trained language model embeddings may infer sensitive attributes such as gender even after explicit indicators are removed. In this paper, we evaluate nine pre-trained embedding models on the synthetic FairCVdb dataset, analyzing the informativeness of their embeddings for applicant scoring and their susceptibility to gender leakage, on both original and gender-scrubbed biographies. We further use a multi-task adversarial learning framework with gradient reversal to predict applicant suitability while suppressing gender information from learned representations. Finally, we use a multi-objective Pareto-front-based model selection to balance predictive utility and fairness. Our experimental results show that explicit gender scrubbing substantially reduces but does not eliminate gender leakage, while adversarial learning improves fairness mainly on original biographies and acts as a complementary strategy rather than a substitute for text-level debiasing.

We present AutoJourn, a demonstration system for multi-perspective news generation and bias-aware evaluation using large language models (LLMs). The system tackles three core challenges in responsible automated journalism: extracting diverse perspectives from unstructured social media discussions, generating summaries that preserve viewpoint diversity, and detecting or mitigating bias in AI-generated news. The pipeline integrates advanced prompt engineering with optional retrieval augmentation to produce semantically diverse perspective sets, a multi-perspective summarisation module that merges conflicting viewpoints into balanced summaries, and a bias analysis suite supporting sentence-level bias detection and type classification in the generated news article, and automatic neutralisation. Users can inspect perspective clusters, compare stance-specific summaries, generate news articles, and apply bias-aware rewrites directly in the interface. We evaluate each component with intrinsic metrics -- semantic diversity, summary quality, and bias reduction and show improvements over strong baselines while maintaining content fidelity. A live, publicly accessible demo accompanies the paper to facilitate reproducibility and further research on socially responsible automated journalism.

As large language model agents gain access to increasingly large skill libraries, retrieving the right skill becomes critical to reliable capability selection and execution. Existing retrievers often treat skill contents as ordinary documents, overlooking their highly regular structure: shared descriptive patterns recur across many skills while providing little evidence for distinguishing the required capability. We show that this shared descriptive background is reflected in dense relevance scores, induces a pronounced energy gap between queries and skill documents, and obscures discriminative signals, especially for structurally similar hard negatives. Based on this observation, we propose SkillSight, a training-free retrieval framework that calibrates shared background in both semantic and lexical spaces. Semantic Background Calibration estimates a background subspace from generic tokens identified by IDF, reducing similarity induced by shared descriptive patterns, while Lexical Evidence Calibration downweights shared background tokens to recover discriminative token-level evidence. Experiments on SRA-Bench and SkillBench-Supp demonstrate consistent improvements across retrieval metrics, with SkillSight improving Recall@10 by up to 20.21 percentage points over the original dense retriever. It is up to 1,248 times faster than the Dense + Reranker baseline. In end-to-end evaluation, SkillSight achieves the best overall performance across three agent models and outperforms LLM Selection by up to 4.97 percentage points. These results identify shared descriptive background as a source of ranking interference in skill retrieval and demonstrate that calibrating it enables accurate and efficient skill selection without additional training. Our code can be found at https://github.com/xiaojinying/SkillSight

Transformers with relative positional encodings often extrapolate to sequences longer than those seen during training, whereas transformers with learned absolute encodings typically do not. This is a robust empirical regularity, and the explanations offered for it so far are chiefly about expressivity, that is, about whether a length-generalizing solution exists. We give an optimization explanation. On a minimal fixed-offset retrieval task that isolates positional selection, the gap is governed by the implicit bias of the trained attention head: among the many solutions that fit short sequences, which one gradient descent actually selects. We prove that rotary encodings make the attention logit a function of relative offset alone, an exact equivariance, so whatever selection rule is learned at training lengths is reproduced verbatim at every longer length. Learned absolute encodings instead leave out-of-range positions unconstrained, and the trained head pins to a fixed absolute position inside the training range. We characterize the learned rotary rule as a low-rank ``carrier'' kernel aligned with the target offset, and we derive the resulting graceful accuracy decay as an attention-dilution law; both predictions are confirmed across seeds and offsets. A linear-attention control shows the mechanism is specific to softmax: without normalization, training selects a min-norm interpolant that does not extrapolate. The phenomenon, the equivariance, and the carrier all transfer to a multi-layer, multi-head transformer trained on a full-sequence length-generalization task. The account connects the implicit bias of attention, implicit bias for extrapolation in recurrent models, and the learning side of the RASP-L conjecture.

Purpose: To develop and validate a deep learning ensemble for estimating adult sex, age, height, and weight from coronal digitally reconstructed radiographs (DRRs) generated from diagnostic CT. Materials and Methods: This retrospective study included 128,621 CT examinations from 80,004 adults at nine institutions in Japan. Three multitask models-ConvNeXt-Base, ViT-Base/16, and MaxViT-Base-were fine-tuned using coronal DRRs and combined by weighted averaging. Data were split by institution into training (114,147 examinations; seven institutions), tuning (4,305; one institution), and test (10,169; one institution) sets; generalizability was assessed on two non-Japanese datasets. Accuracy and mean absolute error (MAE) were used to evaluate sex classification and age, height, and weight regression, respectively. Body surface area (BSA)-corrected heart and liver volume trends were compared using true versus estimated height and weight. Results: In the test set (median age, 69.9 years; 4,899 of 10,169 [48.2%] male), overall sex-classification accuracy was 0.997 (95% CI, 0.996-0.998), and MAEs were 3.57 years (3.51-3.63), 2.59 cm (2.54-2.64), and 3.40 kg (3.34-3.47) for age, height, and weight, respectively. In examinations covering the chest through pelvis, accuracy was 1.000, and MAEs were 3.15 years, 2.28 cm, and 3.18 kg, respectively. BSA calculated from estimated values reproduced age-related heart and liver volume trends obtained using true values. On non-Japanese datasets, height error increased but was reduced by continued fine-tuning. Conclusion: The ensemble estimated adult sex, age, height, and weight from CT-derived DRRs, with generally lower errors in examinations with broader anatomical coverage.

Overparameterized models often have continuous parameter symmetries, so different parameters define the same predictor. We show that PAC--Bayesian analysis should be performed on the quotient predictor space: pushing a prior and posterior to the quotient preserves the empirical and population Gibbs risks while removing the nonnegative KL contribution caused solely by how the two distributions differ among parameterizations of the same predictor. Quotienting alone does not determine which prior to use. We construct a canonical choice of one parameterization for each predictor and account for the geometric volume of its equivalent parameterizations. This transforms a neutral reference prior into a data-independent prior that reflects the model's implicit bias. It approximates the ideal but inadmissible posterior-matched prior, which would minimize the KL term by depending on the training data. The resulting certificate is tighter exactly when this geometry-induced prior has smaller KL divergence from the learned quotient posterior than the neutral prior. We test this prediction in Fourier regression with a Hadamard parameterization and in Query-Key attention, using ordinary SGD without an explicit regularizer. The implicit-bias prior reduces the mean quotient-space KL by $40.69\%$ and the mean PAC--Bayes certificate by $21.40\%$ in the Fourier-Hadamard experiment. The smaller, prior-scale-dependent improvement in Query-Key attention confirms the predicted conditional nature of the effect.

The proliferation of internet memes has introduced new complexities to automated content moderation, particularly in detecting misogyny. Memes often rely on a semantic clash between visual and textual modalities, where hateful intent is implicit and culturally grounded. This paper presents GeoMVC (Geometric Interaction and Multi-View Consensus), developed for the CC-MMD Grand Challenge at ICMI 2026. To address the limitations of static feature concatenation, a Geometric Interaction Layer is proposed that models cross-modal alignment via Hadamard products and cosine similarity between frozen visual and textual embeddings. We further mitigate distribution shifts caused by noisy OCR and code-mixed transliteration through a Multi-View Consensus strategy, aggregating predictions across raw, length-filtered, and English-translated text views. The system achieved Rank 2 in the Malayalam partition (Macro F1: 0.892) and Rank 3 in the Chinese partition (Macro F1: 0.895) on Task A, while securing Rank 5 in the Tamil partition (Macro F1: 0.521). A detailed error analysis on the development partition highlights open challenges in modeling localized transliteration and code-mixed sarcasm across Dravidian and Chinese cultural contexts.

Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.

Many Bangla words are at once personal names and culturally loaded common nouns, "Maya" is both a girl's name and a word for affectionate compassion. Choosing the right reading demands cultural knowledge that is scarce in the pretraining data of modern language models. We introduce Culturally Entangled Homograph (CEH) disambiguation and build a Bangla benchmark of 1,516 expert-verified sentences (3,032 labelled occurrences) in which one word appears twice with two distinct readings, each labelled with a culturally grounded category and an explanation of the reasoning behind it. Across open- and closed-source models, we find a systematic dominant-meaning bias: models default to the common-noun sense and overlook the name. A Bangla-specific model fails under every prompting regime we test, showing that language-specific pretraining alone does not confer cultural grounding. We further show that contrastive chain-of-thought prompting can sharply reduce this bias without training, and that distilling cultural explanations teaches small (1-3B) models to reason toward the correct reading rather than memorise labels, cutting dominant-meaning bias from as high as 100% to under 5% and turning the failed Bangla-specific model into our strongest system. Dataset and code are available at https://github.com/ashuvo25/BanglaCEH.

Online encyclopedias shape political opinion and, through it, democratic discourse. In late 2025, Grokipedia was released, an encyclopedia written entirely by the LLM Grok. One motivation behind the project was to provide an unbiased alternative to Wikipedia, which has faced accusations of "left-wing" and "liberal" bias. But does an encyclopedia written by an LLM deliver greater neutrality, or does it simply embed a different ideology? We conduct a large-scale political bias study on Grokipedia and Wikipedia, analysing 1,394 article pairs describing members of government for neutrality along nine expert-coded ideology dimensions employing four LLM judges, Grok, Claude, Mistral, and DeepSeek. As the LLMs could themselves be biased, we also investigate patterns in their judgments. We find all LLM-judges, including Grok, to rate Grokipedia less neutral than Wikipedia. Both encyclopedias are rated as portraying politicians favourably overall, but towards different ideological groups. Grokipedia particularly favours economically right-wing politicians and penalises socially liberal ones, while Wikipedia is rated as favourably biased towards the latter.

Rare concept generation focuses on synthesizing customized images conditioned on text prompts that describe objects with unusual attributes. Previous works failed to align the generated images with rare concepts, resulting in incorrect attribute rendering or inconsistent composition of concepts. Such failures, as we observed, stem from the inherent common knowledge bias in the training stage of diffusion models, where objects are strongly associated with their common attributes, making it difficult to break these associations when generating rare concepts. To address such challenges, in this paper, we propose a novel Counterfactual Inference-based Diffusion approach, dubbed CI-Diff. CI-Diff blocks the interference of the model's inherent common knowledge bias and utilizes the Natural Direct Effect to capture the independent influence of the text prompt of rare concepts on image generation so that decoupling the unusual attributes from the rare concepts. To this end, we reformulate the classifier-free guidance mechanism to highlight the atypical attributes. To the best of our knowledge, we are the first to introduce causal inference into the rare concept generation task. Extensive experiments on the RareBench benchmark validate the superiority of CI-Diff over state-of-the-art diffusion models. Our code can be accessed from https://github.com/200204jzy/CI-Diff.

This study examines the cross-prompt generalization and first-language (L1) scoring effects of a LoRA-adapted open-weight large language model (Gemma-3-27B-it) applied to automated essay scoring. Using the identical model and inference configuration reported in "AiAWE: An Open-Source LLM Automated Writing Evaluation System Using LoRA-Adapted Instruction-Tuned Models" (Gayed, 2026), which was fine-tuned on 480 argumentative essays from two prompts, we evaluate scoring accuracy on the full TOEFL11 corpus: 12,100 essays written by test-takers from 11 first-language backgrounds across eight prompts, none of which were seen during training. The model's raw scores (0.5-5.0) are mapped to the same three proficiency bands (low, medium, high) used by ETS, enabling direct comparison. The model achieved an overall band agreement of 77.79% and a quadratic weighted kappa of 0.702, with adjacent-band agreement of 99.98%. Accuracy was stable across all eight unseen prompts, with no advantage for prompts thematically related to the training data, indicating robust cross-prompt generalization. However, the model exhibited a systematic, L1-linked scoring offset. Within every proficiency band, essays from European-language backgrounds received consistently higher scores than essays from East-Asian-language backgrounds, a pattern not attributable to the composition of the fine-tuning data. This is the first large-scale L1 fairness analysis of a fine-tuned open-weight LLM for automated essay scoring.

Robots are increasingly integrated into everyday contexts, including museums, where they can both entertain and educate visitors. To enhance visitor experience and engagement, we present a novel mixed-agent tour guide system that combines a physical robot with a projected virtual agent that actively participates in the tour through conversation and interaction, achieving the interaction richness of two mobile agents from a single platform. We validate the system through a within-subjects study with 30 participants to assess engagement, quality of experience, and learning performance. Participants experienced different conversational styles and agent configurations, and data were collected via surveys, behavioral sensors, and interviews. Results showed that engagement and quality of experience remained consistent across conditions. Learning performance revealed a significant gender-moderated difference: the mixed-agent conditions improved learning performance for female participants. This suggests that the proposed dyadic conversational style in this paper influenced learning performance differently by gender. Nonetheless, in interviews, participants reported a greater preference for mixed-agent teams regardless of gender, citing interaction as a key factor in their experience.

Music aesthetics scoring plays a critical role in applications such as dataset curation, generative model evaluation, and reward modeling for music generation. Recent approaches rely on deep neural networks trained on human-annotated ratings, but these models may exploit spurious correlations rather than capturing perceptually meaningful aesthetics. In this work, we identify a previously underexplored failure mode in music evaluation models: genre-induced shortcut learning. Through a systematic analysis of SongEval, we show that biases in training data lead to strong correlations between genre-related features and predicted scores, causing the model to use them as a proxy for aesthetics. This results in systematic overestimation of pop music and undervaluation of high-quality samples from other genres, leading to predictions that are inconsistent with human preferences. To address this issue, we propose a training objective that jointly reweights hard samples and regularizes group-level performance, encouraging the model to learn genre-invariant representations of musicality. Experimental results demonstrate that our method reduces genre-dependent bias and improves alignment with human preferences, as reflected by gains in both cross-genre and within-genre preference alignment.

Clicks on homepage marketing blocks are driven by a dual-mechanism of content interest and access habits. However, habitual clicks often create Pseudo-Positives in marketing slots, where position advantage masks mediocre content quality, leading to biased recommendation ecosystems. We propose a framework called Orthogonal Disentanglement of Access habits (OrDA) to purify interest signals. OrDA utilizes a dual-tower structure with a gated allocation layer to adaptively route features and minimize interference. To ensure rigorous separation, we employ orthogonal regularization to constrain the latent interest and habit manifolds to be geometrically perpendicular. OrDA performs causal intervention (do-calculus) during inference to rank items solely by purified interest scores. Empirical online evaluations on large-scale datasets demonstrate that OrDA effectively eliminates access-habit bias, outperforming state-of-the-art methods in predictive accuracy. Online AB test 5.64% shows user click-through rates (UCTR) improvement on the Zhima homepage marketing block, Zhima rent-floor recommendation.

A field can reformulate its computations freely exactly where its demand is stated independently of any incumbent implementation, and finds itself unable to when the incumbent's own output has quietly become the specification. This note offers that observation as a lens on computational reformulation for modern accelerators, where posing a problem in a hardware-friendly form can yield large speed and energy gains, but only if a replacement can be judged at all. Building on the test-oracle problem (Weyuker; Barr et al.), on requirements engineering's notion of implementation bias (Zave and Jackson), and on the roofline performance model, it names the pathology "baseline capture" -- the moment an incumbent stops being evidence that a demand can be met and becomes the definition of meeting it -- and separates two questions that are easily confused: whether a reformulation can be judged (which turns on the existence of an incumbent-independent demand) and whether its discovery can be automated (which turns additionally on the cost of evaluating that demand). Short cases -- shortest-path routing, learnable audio frontends, ZIP-215 for Ed25519 signature validation, CESM-ECT for climate models, and a single-GEMM audio frontend -- illustrate the pattern and the move of "buying a verifier": making a demand explicit, operational, and independent of the incumbent. No component is claimed novel in isolation; the contribution is the synthesis and the single question it makes easy to ask of any reformulation result -- does its acceptance test mention the incumbent's output?

Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest. On schema-dense input streams such as nested JSON, this score acts as a non-stationary filter that disproportionately retains noise: a non-content sink role (delimiters or whitespace) carries an order of magnitude more energy than any content role, and structural KEY tokens are over-retained at roughly 1.8x the rate of the answer-carrying VALUE tokens, collapsing exact-match accuracy from 88% to 0% at a 5% budget as the signal-to-noise ratio of the retained state degrades. A counterfactual experiment establishes that suppressing KEY tokens is the best deployable filter. Our retraining-free, role-conditional allocation over SnapKV's windowed score, governed by a single tuned hyperparameter, closes 63-98% of the H2O gap at sub-20% budgets and, at higher budgets, modestly matches or exceeds full-cache accuracy -- a small, seed-sensitive denoising effect (borderline significant at B=0.50; not distinguishable from zero at B=0.30 over four seeds). A 15 MB linear role probe supplies these labels at negligible inference cost, though matching parser-level downstream accuracy remains open.

Transformer language models are increasingly used as software components, yet biased outputs remain difficult to localize and repair inside the model. Existing fairness testing and repair methods largely operate at the input-output or retraining level, while recent work suggests that bias-related behavior can concentrate in a small set of attention heads. This paper studies whether attention heads can be localized and repaired through a targeted inference-time intervention. We introduce ROBIN, a white-box head-level fairness debugging method that ranks attention heads using sensitivity to fairness probes and removes a small bias subspace from selected head outputs. In a four-model pilot study, ROBIN reduces the measured WinoBias gap across all models while preserving language-modeling quality better than whole-head zeroing. These preliminary results suggest that head-level bias repair should consider not only which heads are selected, but also how selected heads are modified.

Multiple-choice benchmarks that rank candidate completions by conditional log-probability suffer from a length bias: because log-probabilities sum over tokens, longer answers tend to be penalized relative to shorter ones in practice. A common mitigation is to normalize scores by completion length, but we show empirically that this heuristic frequently over-corrects, introducing a bias toward longer answers instead. We first analyze these scoring rules, characterizing when standard and length-normalized accuracy are appropriate and how their length biases depend on the distribution of completion lengths. Motivated by this analysis, we introduce \emph{Bayesian accuracy}, a scoring rule that computes the posterior probability of each candidate under an explicit prior over answer length, thereby removing linear length effects. Bayesian accuracy is a drop-in replacement for likelihood-based multiple-choice evaluation, requires no additional forward passes, and consistently exhibits lower empirical length bias than both standard and length-normalized accuracy across benchmarks and few-shot settings.

We study the gradient flow dynamics of diagonal linear networks for regression tasks under infinitesimal initialization. Extending Theorem 1 from Pesme & Flammarion (2023), we generalize the analysis to both deep diagonal linear networks and a broader class of two-layer diagonal linear networks (as defined in Definition 4.1). Specifically, we demonstrate that the training trajectories of these models can be equivalently characterized by the proposed Algorithm 1. We further prove that this algorithm converges to the solution of a modified $\mathcal{l}_1$ norm minimization problem. As a result, we establish that the implicit bias of both network architectures corresponds to a modified $\mathcal{l}_1$ norm in the regime of infinitesimal initialization. Additionally, we provide insights into the underlying mechanisms governing these dynamics by identifying the Structural Invariant Manifold (SIM) (Zhao et al., 2026) as the key geometric structure that shapes the learning process.

Audio-language embedding models such as CLAP are widely evaluated on matching present sound events, but rarely on negation. We show this affirmation-only evaluation hides a key limitation: these models fail to encode negated sound concepts, mapping affirmative and negated captions to nearly identical representations. To expose this blind spot, we introduce NegEval-Audio, a framework that converts existing datasets into two negation-aware tasks, Retrieval-Neg and Multiple-Choice Negation (MCQ-Neg), to probe whether models distinguish present from absent events. On AudioCaps and Clotho, performance degrades sharply under negation, with negation-type MCQ accuracy falling far below chance, and the failure persists even for a recent multimodal LLM-based embedding model. While a training-free steering method improves MCQ-Neg, it yields marginal gains for Retrieval-Neg. This indicates that affirmation bias is a fundamental flaw in the representation geometry, necessitating explicit negation-aware training objectives.

Accurately assessing personality from text is challenging because traits are latent, context-dependent, and often subtly expressed across long narratives. Large language models (LLMs) offer new opportunities by processing extensive textual contexts, but pretraining of these models can induce latent "personality-like" biases, making single-model inferences inconsistent. We propose a fine-tuned multi-agent framework for detecting OCEAN personality traits, in which sub-agents are conditioned to adopt high, low, or neutral perspectives for each trait through masked language modeling (MLM) and psychometric supervision. A judge LLM aggregates and compares sub-agent outputs to generate final trait predictions, capturing multiple complementary perspectives while mitigating individual model biases. We evaluate the framework on life narrative dataset through quantitative and qualitative experiments, including baselines, ablations, and inference quality analyses. Our approach offers a scalable and interpretable method for text-based personality inference, highlighting the benefits of multi-agent reasoning grounded in psychometric supervision.

Retrieval-Augmented Generation (RAG) has been increasingly adopted to reduce hallucinations and strengthen the factual grounding of large language models (LLMs). While robustness to errors in the retrieval process has been explored, the impact of ideological bias on LLM outputs has been overlooked. For instance, if the retrieved material contains ideological positions, the RAG may transmit, amplify, or suppress such ideological discourses in its outputs. In this study, we address this issue by examining the influence of the RAG framework, comprising ideological discourses, in LLM-generated answers. To this end, we applied Lexical Multidimensional Analysis (LMDA) on a corpus of 1,117 COVID-19 treatment articles, identifying three ideological discourses. This corpus is then used as the external knowledge source for the RAG. We assessed several LLMs by having the models answer ideological questions at different sampling temperatures. The generated texts were assessed semantically and lexically based on their similarities with ideological reference texts. Our findings show that the RAG framework is prone to transferring ideological discourses into LLM responses, with sampling temperature having a measurable impact on the strength of this transfer. Discoursive alignment between generated answers and the reference text is highest at moderate temperatures, where models balance stochasticity with retrieval grounding, and drops at low temperatures, indicating that overly deterministic sampling suppresses discourse transfer.

Human choice behavior, including route choice, exhibits systematic behavioral biases that deviate from the assumptions of full rationality. Cumulative prospect theory (CPT) has been widely recognized as an effective framework for characterizing such behavioral patterns. However, its large-scale application, particularly in simulation and agent-based modeling, critically depends on specifying individual-level CPT parameters, which remain a major bottleneck. Conventional approaches typically rely on surveys and controlled experiments to calibrate CPT parameters, yet these methods are difficult to generalize and often fail to capture the full diversity of human decision-making. To address this challenge, this paper investigates whether large language models (LLMs) can reproduce human behavioral biases in choice-making without explicit specification of prospect-theoretic parameters. Using route choice as a representative scenario, we design a behavioral evaluation framework and systematically compare LLM-generated decisions with established human behavioral patterns predicted by CPT. Experimental results demonstrate that LLMs are capable of reproducing non-rational human choice biases and can exhibit decision behaviors consistent with prospect-theoretic effects under uncertainty. These findings suggest that generative AI models may provide a scalable alternative for modeling human decision processes and offer a promising foundation for next-generation large-scale agent-based simulation and AI-driven behavioral research.

A lot of research attention has been devoted to checking whether large language models (LLMs) are politically biased. This work has largely focused on high-level ideological dimensions, such as left--right or progressive--conservative, and it has been shown that while LLMs are predominantly left and progressive leaning, largely mimicking the biases in the training data, they can be to some extent steered to change their preferences in post-training. In this short note, we check if LLMs have robust stances with regard to major substantive societal issues, on which members of the same ideological camp are often in disagreement, summarised in a novel dataset \textsc{HardChoices}. We show that, faced with this line of questioning, LLMs, both large and small, surprisingly rarely declare neutrality, are often incoherent, and demonstrate a remarkable degree of agreement on issues where they do take stances.

As Large Language Models (LLMs) are increasingly deployed as conversational tutors, they risk institutionalizing systemic inequalities. This study presents a systematic API audit of four LLMs acting as history tutors, evaluating 1,800 responses regarding the 1989 Romanian Revolution across five student personas varying by ethnicity and socio-economic tier. We uncover four interconnected patterns of \emph{epistemic paternalism}: (1)\textbf{Differential Refusal}, where safety-aligned models block 76.7% of educational requests from low-tier students; (2)\textbf{Epistemic Gatekeeping}, evidenced by a 3$\times$ reduction in access to geopolitical complexity (e.g., the contested ``coup theory'') for marginalized learners; (3)\textbf{Agency Theft}, a lexical shift where models like LLaMA produce a 5$\times$ higher victimization-to-politics vocabulary ratio for Roma students compared to elite peers; and (4)\textbf{Elite Hermeneutics}, where AI tutors disproportionately withhold epistemic confidence and justification scores from low-resource demographic profiles. We argue that current safety alignment acts as a paternalistic filter, transforming conversational AI into agents of narrative segregation -- a manifestation of \emph{hermeneutical injustice} in Fricker's~\cite{fricker2007} sense that demands urgent pedagogical auditing.

While Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities, they remain highly susceptible to embedded social biases. Existing bias evaluation protocols predominantly rely on static datasets, which provide only a superficial assessment, as their fixed test cases cannot adaptively evolve to measure the true depth and limits of model vulnerabilities. We introduce DeepBias, an adaptive framework for the in-depth probing of social biases in LVLMs with carefully designed agents. Our approach operates through a dynamic ''generation-evolution-probing'' loop. First, a generative ProposerAgent synthesizes test data and is iteratively updated via Direct Preference Optimization (DPO) based on the target LVLM's responses, exploring model-specific failure modes. Second, an autonomous skill-driven DiggerAgent rewrites each test data across multiple probing turns, adaptively selecting from a curated skill library of deepening and rewriting strategies. At each turn, this process is conditioned on the model's previous response, enabling progressively deeper biases to be exposed. Furthermore, we build a benchmark named DeepBiasBench using our framework. By employing an ensemble of five diverse state-of-the-art LVLMs as anchors, the benchmark captures vulnerabilities shared across architectures. Comprehensive experiments demonstrate the effectiveness of our framework and show that DeepBias provides a challenging benchmark for in-depth bias evaluation, establishing an evolutionary paradigm for LVLM safety assessment.

Raw images inherently suffer from noise due to the stochastic nature of light and sensor hardware imperfections. As real photon counts fall, the ratio of this noise to the signal degrades; consequently, for low-light conditions, robust denoising is especially vital for high-quality results. While recent data-driven methods achieve strong performance, they typically rely on large-scale noisy-clean image pairs that are costly and difficult to collect. Alternatively, parametric noise models can generate synthetic training data, but this necessitates precise camera calibration, which is often impractical for unknown devices. In this work, we propose a camera-agnostic, calibration-free paradigm for low-light raw denoising. We identify that color bias from black-level error is a primary source of performance degradation and causes severe color shifts. To mitigate this, we introduce a bias estimator network that predicts the black-level error as a global feature of the noisy input. We evaluate our approach across the ELD, SID, and LRID datasets, demonstrating superior performance among blind denoisers, particularly in terms of color correction. In many cases, we are competitive with-or can even surpass-methods with stronger supervision. Furthermore, we reveal that the widely used SIDD dataset contains significant color bias in its ground-truth images, which yields unrealistic color reproduction in trained models. We introduce a new ground-truth extraction framework to resolve this issue and provide a benchmark of existing methods on the corrected dataset.

Can a Video Large Language Model (Video-LLM) follow one person through a long video, keeping track of who they are well enough to report, in order, how their outfit changes across a full TV episode? Benchmarks increasingly score this kind of task, and the strongest open-source 7--8B models now reach 37--38% on InfiniBench's global appearance task, which asks exactly that. But does that score come from tracking the named character, or from something easier? We test this with a nine-condition diagnostic protocol applied to three architecturally distinct open-source Video-LLMs, with Gemini2.5Flash as a frontier reference, and find the accuracy does not come from character tracking. When we change the character named in the question to a different cast member, leaving the video and answer options untouched, the models change their answer only 4--31% of the time, so they are largely ignoring who the question asks about. Breaking that test down by the gender of the swapped name shows why: the models react more when the name is changed to a different-gender character than to a same-gender one (a 13--28 point gap), picking up coarse gender cues but unable to tell same-gender individuals apart. This shallow processing surfaces again when we drop the multiple-choice options and ask the same questions open-endedly: open-source accuracy drops 18--25 points, with none of 151 answers fully correct, versus a 12-point drop for Gemini. Further checks rule out the obvious innocent explanations, adding subtitles, using the most informative frames, or doubling the number of frames all leave character tracking unimproved, so the bottleneck is not how much video the model sees but how it ties that video to the person the question names. We release a diagnostic toolkit for auditing what such benchmark scores actually measure.

Audio deepfake detection models determine whether speech is genuine or artificially generated, but high overall accuracy can mask substantial performance disparities across demographic groups. In this work, we investigate gender bias in audio deepfake detection using the ASVspoof5 dataset. We use ASVspoof5 under a controlled custom split designed to isolate gender-composition effects. We train attack-specific models on nine training sets with different gender compositions, ranging from female-only to male-only. We use a ResNet18 classifier with LogSpectrogram and WavLM-Base+ features, and we evaluated six post-hoc threshold calibration methods. Experimental results show that training data composition strongly predicts bias direction, with the underrepresented gender performing worse at test time. WavLM-Base+ features are shown to produce gender performance gaps 3.0 to 4.3 times larger than LogSpectrogram under identical training conditions, and balanced training is found to reduce LogSpectrogram bias but leave WavLM bias largely intact. Moreover, all six calibration strategies, including Oracle calibration with full test-set label access, leave the Equal Error Rate gap unchanged at 1.317 pp, confirming that threshold adjustment cannot correct underlying score distribution disparities. Overall, these findings suggest that gender fairness in audio deepfake detection must be addressed at training time, as post-hoc methods can only partially mitigate the resulting disparities

Algorithmic fairness methods are increasingly used to identify and mitigate bias in machine learning models, yet most approaches are evaluated in isolation and along single demographic axes. This limits practical guidance for selecting fairness strategies, where disparities may arise across intersectional subgroups and across multiple stages of the modeling lifecycle. This work presents FairSelect, a toolkit for systematically evaluating fairness mitigation strategies applied individually and in combination across preprocessing, inprocessing, and postprocessing stages. FairSelect supports multiple model architectures, intersectional subgroup evaluation, and comparison of fairness utility tradeoffs across baseline, single method, and multi level configurations. The framework was validated using synthetic clinical datasets designed to represent specific bias mechanisms and a real-world replication of two-year stroke risk prediction among patients with atrial fibrillation. Synthetic experiments showed that targeted fairness methods generally reduced intended subgroup disparities, while combined strategies produced larger average fairness improvements with modest utility tradeoffs. In the clinical prediction task, mitigation effects were highly variable, with some combinations improving both fairness and predictive performance while others were ineffective or counterproductive. These findings demonstrate that fairness interventions interact in nonadditive and context dependent ways. FairSelect provides a practical framework for systematically identifying fairness strategies that improve subgroup equity while preserving model performance in clinical machine learning.

Text-to-image (T2I) models have been shown to exhibit social biases. Prior work has mainly focused on gender, skin tone, and cultural representation within restricted occupational associations, and emerging benchmarks increasingly incorporate these dimensions. However, disability remains systematically underexplored. Current evaluation practices often fail to align with sociologically grounded definitions of stereotyping, limiting principled assessment of representational harms toward people with disabilities (PWD). To address this, we introduce INCLUDE-BENCH, the first large-scale benchmark for evaluating disability-related bias in T2I models. INCLUDE-BENCH comprises 119K generated images based on prompt design across multiple bias dimensions and both static and dynamic contexts. We evaluate 15 open-source and 2 closed-source models. Our key findings reveal that: (1) mobility-impaired and default disability prompts predominantly yield wheelchair depictions across all models; (2) disability-conditioned generations consistently exhibit less diversity; (3) stereotypical portrayals demonstrate stronger disability-text alignment; and (4) we introduce the Stereotype Content Model (SCM) Score, demonstrating that T2I models reflect real-world stereotypical associations.

Vision-language alignment, the stage that bridges pretrained vision encoders and large language models, is widely treated as a form of pretraining requiring full-parameter updates. We challenge this view and investigate what happens when low-rank adaptation is applied to the LLM during this stage instead. We find that low-rank alignment not only reduces computational costs but also outperforms full-parameter alignment on most benchmarks. To understand this phenomenon, we systematically characterize the implicit biases introduced by low-rank adaptation during alignment. Empirically, we find that low-rank alignment shifts model behavior from hallucinatory to conservative and preserves per-token linear separability of visual features that full-parameter alignment disrupts, a phenomenon we term LS-curse. Geometrically, low rank aligned models exhibit more homogeneous and structurally stable visual representations, maintaining modality-specific knowledge rather than prematurely fusing entity-level semantics. Theoretically, we establish two theorems showing that low-rank alignment induces preferences for parameter subspaces with flat gradients and feature subspaces robust to perturbations, providing a principled explanation for the observed structure-preserving behavior. Extensive experiments cover ablation over 100 alignment configurations, three families of low-rank operators, and various rank, encoder, and other settings.

Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.