Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments

• URL: http://arxiv.org/abs/2604.02669v1
• Date: Fri, 03 Apr 2026 03:03:21 GMT
• Title: Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments
• Abstract summary: Single-task benchmarks miss this because they capture only one slice of a model's bias profile.<n>We introduce a hierarchical taxonomy covering 9 bias types, including under-studied axes like caste, linguistic, and geographic bias.<n>We Auditing 7 commercial and open-weight LLMs with textasciitilde45K prompts, we find three systematic patterns.
• Score: 0.19703625025720697
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: How biased is a language model? The answer depends on how you ask. A model that refuses to choose between castes for a leadership role will, in a fill-in-the-blank task, reliably associate upper castes with purity and lower castes with lack of hygiene. Single-task benchmarks miss this because they capture only one slice of a model's bias profile. We introduce a hierarchical taxonomy covering 9 bias types, including under-studied axes like caste, linguistic, and geographic bias, operationalized through 7 evaluation tasks that span explicit decision-making to implicit association. Auditing 7 commercial and open-weight LLMs with \textasciitilde45K prompts, we find three systematic patterns. First, bias is task-dependent: models counter stereotypes on explicit probes but reproduce them on implicit ones, with Stereotype Score divergences up to 0.43 between task types for the same model and identity groups. Second, safety alignment is asymmetric: models refuse to assign negative traits to marginalized groups, but freely associate positive traits with privileged ones. Third, under-studied bias axes show the strongest stereotyping across all models, suggesting alignment effort tracks benchmark coverage rather than harm severity. These results demonstrate that single-benchmark audits systematically mischaracterize LLM bias and that current alignment practices mask representational harm rather than mitigating it.

Related papers

• How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs? [53.58941416780391]
  We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model.<n>We extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out transfer, and causal intervention.
  arXiv  Detail & Related papers  (2026-07-20T16:10:06Z)
• Dual-Branch Cross-Projection Debiasing through Diffusion-based Disentanglement [66.67463557274358]
  Foundation models trained on biased datasets often rely on spurious correlations between target labels and non-causal attributes.<n>We introduce Confidence-guided Concept Mining (CBCM) to identify reliable spurious attributes without attribute annotations.<n>We propose Dual-branch Cross-projection Debiasing (DCD), a prompt-tuning framework that separates target and spurious representations into two branches.
  arXiv  Detail & Related papers  (2026-06-23T05:28:47Z)
• Moral Sensitivity in LLMs: A Tiered Evaluation of Contextual Bias via Behavioral Profiling and Mechanistic Interpretability [22.32075837181307]
  Large language models (LLMs) are increasingly deployed in settings that require nuanced ethical reasoning.<n>We introduce the Moral Sensitivity Index (MSI), a metric that quantifies the probability of biased output.<n>We select criminal-bias scenarios, which produced the highest MSI scores across models.
  arXiv  Detail & Related papers  (2026-05-04T23:12:32Z)
• Addressing Stereotypes in Large Language Models: A Critical Examination and Mitigation [0.0]
  Large Language models (LLMs) have gained popularity in recent years with the advancement of Natural Language Processing (NLP)<n>This study inspects and highlights the need to address biases in LLMs amid growing generative Artificial Intelligence (AI)<n>We utilize bias-specific benchmarks such StereoSet and CrowSPairs to evaluate the existence of various biases in many different generative models such as BERT, GPT 3.5, and ADA.
  arXiv  Detail & Related papers  (2025-11-18T05:43:34Z)
• Robustly Improving LLM Fairness in Realistic Settings via Interpretability [0.16843915833103415]
  Anti-bias prompts fail when realistic contextual details are introduced.<n>We find that adding realistic context such as company names, culture descriptions from public careers pages, and selective hiring constraints induces significant racial and gender biases.<n>Our internal bias mitigation identifies race and gender-correlated directions and applies affine concept editing at inference time.
  arXiv  Detail & Related papers  (2025-06-12T17:34:38Z)
• Surface Fairness, Deep Bias: A Comparative Study of Bias in Language Models [45.41676783204022]
  We investigate various proxy measures of bias in large language models (LLMs)<n>We find that evaluating models with pre-prompted personae on a multi-subject benchmark (MMLU) leads to negligible and mostly random differences in scores.<n>With the recent trend for LLM assistant memory and personalization, these problems open up from a different angle.
  arXiv  Detail & Related papers  (2025-06-12T08:47:40Z)
• Implicit Bias in LLMs: A Survey [2.07180164747172]
  This paper provides a comprehensive review of the existing literature on implicit bias in Large language models.<n>We begin by introducing key concepts, theories and methods related to implicit bias in psychology.<n>We categorize detection methods into three primary approaches: word association, task-oriented text generation and decision-making.
  arXiv  Detail & Related papers  (2025-03-04T16:49:37Z)
• The Impact of Unstated Norms in Bias Analysis of Language Models [0.03495246564946556]
  Counterfactual bias evaluation is a widely used approach to quantifying bias.<n>We find that template-based probes can lead to unrealistic bias measurements.
  arXiv  Detail & Related papers  (2024-04-04T14:24:06Z)
• Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes [73.12947922129261]
  We leverage the zero-shot capabilities of large language models to reduce stereotyping. We show that self-debiasing can significantly reduce the degree of stereotyping across nine different social groups. We hope this work opens inquiry into other zero-shot techniques for bias mitigation.
  arXiv  Detail & Related papers  (2024-02-03T01:40:11Z)
• What Do Llamas Really Think? Revealing Preference Biases in Language Model Representations [62.91799637259657]
  Do large language models (LLMs) exhibit sociodemographic biases, even when they decline to respond? We study this research question by probing contextualized embeddings and exploring whether this bias is encoded in its latent representations. We propose a logistic Bradley-Terry probe which predicts word pair preferences of LLMs from the words' hidden vectors.
  arXiv  Detail & Related papers  (2023-11-30T18:53:13Z)
• Evaluating the Fairness of Discriminative Foundation Models in Computer Vision [51.176061115977774]
  We propose a novel taxonomy for bias evaluation of discriminative foundation models, such as Contrastive Language-Pretraining (CLIP) We then systematically evaluate existing methods for mitigating bias in these models with respect to our taxonomy. Specifically, we evaluate OpenAI's CLIP and OpenCLIP models for key applications, such as zero-shot classification, image retrieval and image captioning.
  arXiv  Detail & Related papers  (2023-10-18T10:32:39Z)
• Causally Testing Gender Bias in LLMs: A Case Study on Occupational Bias [33.99768156365231]
  We introduce a causal formulation for bias measurement in generative language models.<n>We propose a benchmark called OccuGender, with a bias-measuring procedure to investigate occupational gender bias.<n>The results show that these models exhibit substantial occupational gender bias.
  arXiv  Detail & Related papers  (2022-12-20T22:41:24Z)
• UnQovering Stereotyping Biases via Underspecified Questions [68.81749777034409]
  We present UNQOVER, a framework to probe and quantify biases through underspecified questions. We show that a naive use of model scores can lead to incorrect bias estimates due to two forms of reasoning errors. We use this metric to analyze four important classes of stereotypes: gender, nationality, ethnicity, and religion.
  arXiv  Detail & Related papers  (2020-10-06T01:49:52Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.