Tree-based models such as random forests tend to assign higher importance scores to continuous predictors than to categorical ones. The authors analyze the source of this bias theoretically and suggest a straightforward fix: injecting a small amount of noise into each categorical predictor. The work targets mixed data settings where such scoring imbalance is common.
A new preprint on arXiv studies why a teacher model's bias can still pass to a student model even when the training data has had all overt mentions of that bias removed. The authors report that no data-level defense tested so far reliably detects or eliminates this residual, or "phantom," transfer. The work frames the phenomenon as a dataset-level problem rooted in subtle statistical traces rather than explicit labels.
A new preprint looks at how generative AI writing tools and the language models behind them may shape the political information voters encounter ahead of Sweden's 2026 election. The authors argue that as these assistants become a common way to gather information, their tendency to favor certain issues or viewpoints deserves closer scrutiny. The work adds to a growing body of research on how model behavior can sway user opinions.
A research team called VANGUARD submitted a framework for the EXIST 2026 Task 2 challenge that treats sexism detection as a subjective task shaped by individual perspective. The approach combines multimodal analysis with psychological and demographic conditioning, including biometric signals, to better model how different people judge the same content. The work appears in two arXiv cross-listings under cs.AI and cs.LG.
A research paper examines how fairness and bias behave in large language models when several sensitive attributes, such as gender and ethnicity, are considered together rather than one at a time. The authors run a systematic evaluation across six LLMs to assess intersectional fairness, a setting relevant to socially sensitive deployments. The work highlights gaps in how current models handle overlapping demographic characteristics.
A new arXiv paper examines how people from different backgrounds can jointly agree on where the acceptable limits of AI use in policing should lie. The authors note that AI tools are spreading through UK police forces and elsewhere, and that racial bias is a documented problem, yet communities most likely to be affected are seldom consulted about deployments. The work reports findings from a deliberation process that brought affected community representatives together with other stakeholders.
A new arXiv paper examines what happens when human corrections are applied only to identified editable spans of a model's output. The authors decompose the localized gradient into edited and untouched portions at a fixed checkpoint, showing that selective feedback channels can amplify relative selection bias. They also study gradient geometry, target mismatch, and importance weighting as factors in this effect.
A revised arXiv preprint examines disparate impact as a fairness criterion for synthetic data generation, asking whether generated records deliver equal utility across sensitive demographic groups. The authors position their work as a departure from prior fair synthetic-data research, which they say addresses related but distinct fairness goals. The paper is a research contribution and does not announce any released model or tool.
A new arXiv paper studies how contextual bias affects automated code review systems built on large language models, which are increasingly used both as interactive assistants and as autonomous agents in CI/CD pipelines. The authors measure this bias and explore ways it could be exploited, framing the work around the reliability of LLM-driven security review in real development workflows.
A new arXiv preprint examines using counterfactual image generation to augment training data for medical image analysis. The authors argue that biased datasets produce biased models with limited clinical usefulness, and that synthetic counterfactual images can help offset those biases. The work is announced as a new submission in the cs.LG category.
A new arXiv paper argues that clinical language models are frequently assessed on retrospective patient records that already contain the eventual diagnosis, treatment response and outcome. Because those records expose information a real prospective decision-maker would not have, such benchmarks may reward models for exploiting future data instead of genuine reasoning. The work examines how this exposure shapes model judgments in clinical temporal tasks.
A new arXiv paper introduces FairFund-Bench, a benchmark for measuring how large language models distribute scarce resources and whether those allocations vary by race, gender, or similar traits. The authors note that prior audits of LLM bias have yielded conflicting findings, and position their benchmark as a way to standardize such evaluations. The work targets fairness in settings where models take part in allocating limited goods or funds.
A new arXiv paper introduces DiSCo, a framework that tests whether large language models lean toward particular cultures when answering everyday questions grounded in cultural context. Instead of scoring single responses, it analyzes full output distributions and steers model behavior to quantify underlying cultural priors. The authors argue the approach helps developers detect bias that undermines localization and equitable global deployment.
An updated arXiv paper introduces BTBR, a framework grounded in Bayesian theory and probabilistic-fuzzy methods designed to reduce hidden bias in large language models. The authors argue that such bias can remain dormant under standard prompts yet emerge when the model is guided toward particular demographic personas. The listing reflects a revised version (v2) cross-posted across arXiv categories.
An arXiv paper in cs.AI introduces a technique for auditing bias in large language models by analyzing internal hidden states instead of relying on generated outputs. By comparing a model's representations against those of a reference model using relative representations, the approach aims to detect internal bias shifts that output-based benchmarks or judge models could miss. The authors frame it as a cheaper alternative to benchmark-heavy or judge-dependent auditing pipelines.
A new arXiv paper cross-listed in AI and machine learning introduces a statistical framework for examining bias in generalized zero-shot learning, where models must recognize classes that never appeared in training. The authors ground the analysis in handwriting recognition, a setting where skewed distributions can disproportionately affect underrepresented groups. The work aims to extend bias measurement beyond the relatively narrow conditions covered by traditional GZSL methods.
Researchers have released a study disentangling two distinct gender signals that speech-to-speech models can pick up: the acoustic characteristics of a speaker's voice and gender-related cues embedded in what is being said. This distinction matters for applications like dubbing, translation, and voice agents, where an ideal system should preserve how a speaker actually sounds rather than defaulting to stereotyped content. The work offers a framework for auditing whether these models rely on voice or on content-based assumptions when producing gendered output.
A systematic evaluation on arXiv examines whether large language models give consistent predictions when used as ordinal classifiers. The researchers ran controlled experiments showing that semantically equivalent changes to prompt organization, such as the ordering of labels and demonstrations, can shift model outputs. The findings highlight reliability concerns for deploying LLMs in ranking and rating tasks.
A new paper examines whether and how a defendant's courtroom statement shifts the verdicts of large language models simulating jurors in common-law trials. The authors focus on persuasion effects and potential bias in these simulated decisions, a setting where LLM-based modeling of human judgment had not previously been studied.
Large language models often converge on the same answer even when many plausible alternatives exist, a pattern prior work has labeled as bias. A new arXiv paper proposes separating this concentration into stable model preferences versus responses that depend on a specific prompt. The framework aims to clarify when repeated answer selection reflects genuine bias rather than shallow sensitivity to prompt wording.
New arXiv research investigates AI models that detect early signs of dementia and mild cognitive impairment from conversational speech, focusing on multilingual English speakers in the UK. The authors report that such screening tools show a false positive bias, disproportionately flagging multilingual speakers compared with monolingual ones. The findings suggest speech-based cognitive screening may disadvantage linguistically diverse populations unless corrected.
A new arXiv paper examines whether large language models can pick up on emotional nuance conveyed through textual framing, not just surface-level bias. The authors test model alignment across different sociodemographic groups to see how framing choices affect responses. The work positions framing comprehension as a distinct alignment concern beyond conventional bias evaluation.
A new arXiv study examines how large language models reach stigmatizing conclusions about people with mental health conditions, rather than only scoring their final outputs. The authors analyze model reasoning steps to locate where such bias emerges during generation. The work targets evaluations of LLMs proposed for mental health uses, where prior research has documented stigmatizing responses.
A new arXiv paper argues that commercial text-to-image services quietly rewrite user prompts before image generation, a step people usually cannot see or turn off. Prior cultural-bias audits, the authors note, only look at finished images and treat the whole process as one pipeline, which hides where the bias enters. The work calls for examining this hidden revision stage separately.
A new arXiv paper argues that a language model's political leanings are better described as a probability distribution that shifts with the prompt and surrounding context rather than a single stable viewpoint. The authors report empirical tests across nine current LLMs to support this framing of ideology as conditional on context. The work is positioned as a measurement approach for studying political behavior in models.