Data Matters Most: Auditing Social Bias in Contrastive Vision Language Models

• URL: http://arxiv.org/abs/2501.13223v6
• Date: Sun, 26 Oct 2025 20:51:41 GMT
• Title: Data Matters Most: Auditing Social Bias in Contrastive Vision Language Models
• Authors: Zahraa Al Sahili, Ioannis Patras, Matthew Purver,
• Abstract summary: We disentangle three design factors -- model size, training-data scale, and training-data source -- by comparing CLIP and OpenCLIP.<n>We also evaluate three post-hoc, test-time debiasing strategies -- Bias Prompts, Prompt Array, and SANER.
• Score: 28.944990804599893
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Vision-language models (VLMs) deliver strong zero-shot recognition but frequently inherit social biases from their training data. We systematically disentangle three design factors -- model size, training-data scale, and training-data source -- by comparing CLIP and OpenCLIP, two models that share an identical contrastive objective yet differ in encoder width and in the image-text corpora on which they are pre-trained (400M proprietary pairs vs. 400M/2B LAION). Across balanced face-analysis benchmarks, enlarging the encoder reduces gender skew in CLIP but amplifies both gender and racial skew in OpenCLIP; increasing the LAION corpus from 400M to 2B further increases OpenCLIP bias. At matched model and data budgets, substituting proprietary data with LAION improves gender fairness while increasing racial skew, underscoring data source as the primary driver of bias patterns. We also evaluate three post-hoc, test-time debiasing strategies -- Bias Prompts, Prompt Array, and SANER. Debiasing reduces but does not eliminate harm, and its effectiveness is source- and size-dependent: Bias Prompts most effectively reduce gender skew in CLIP at smaller model sizes, whereas Prompt Array and SANER more reliably reduce racial skew in OpenCLIP; scaling LAION reconfigures which method is most fair. Taken together, these findings challenge the assumption that bigger models or datasets are automatically fairer and foreground training data source as the key determinant of both bias and mitigation efficacy. We release code and evaluation scripts to enable transparent, reproducible auditing of future VLMs.

Related papers

• Freeze and Reveal: Exposing Modality Bias in Vision-Language Models [5.368169150936655]
  Vision Language Models achieve impressive multi-modal performance but often inherit gender biases from their training data.<n>In this work, we dissect the contributions of vision and text backbones to these biases by applying targeted debiasing.<n>Our work enables more targeted and effective bias mitigation strategies in future multi-modal systems.
  arXiv  Detail & Related papers  (2025-08-10T17:08:10Z)
• Muddit: Liberating Generation Beyond Text-to-Image with a Unified Discrete Diffusion Model [87.23753533733046]
  We introduce Muddit, a unified discrete diffusion transformer that enables fast and parallel generation across both text and image modalities.<n>Unlike prior unified diffusion models trained from scratch, Muddit integrates strong visual priors from a pretrained text-to-image backbone with a lightweight text decoder.
  arXiv  Detail & Related papers  (2025-05-29T16:15:48Z)
• Weak Supervision Dynamic KL-Weighted Diffusion Models Guided by Large Language Models [0.0]
  We present a novel method for improving text-to-image generation by combining Large Language Models with diffusion models.<n>Our approach incorporates semantic understanding from pre-trained LLMs to guide the generation process.<n>Our method significantly improves both the visual quality and alignment of generated images with text descriptions.
  arXiv  Detail & Related papers  (2025-02-02T15:43:13Z)
• On the Fairness, Diversity and Reliability of Text-to-Image Generative Models [68.62012304574012]
  multimodal generative models have sparked critical discussions on their reliability, fairness and potential for misuse.<n>We propose an evaluation framework to assess model reliability by analyzing responses to global and local perturbations in the embedding space.<n>Our method lays the groundwork for detecting unreliable, bias-injected models and tracing the provenance of embedded biases.
  arXiv  Detail & Related papers  (2024-11-21T09:46:55Z)
• debiaSAE: Benchmarking and Mitigating Vision-Language Model Bias [1.3995965887921709]
  We analyze demographic biases across five models and six datasets.<n>Portrait datasets like UTKFace and CelebA are the best tools for bias detection.<n>Our debiasing method improves fairness, gaining 5-15 points in performance over the baseline.
  arXiv  Detail & Related papers  (2024-10-17T02:03:27Z)
• Collapsed Language Models Promote Fairness [88.48232731113306]
  We find that debiased language models exhibit collapsed alignment between token representations and word embeddings. We design a principled fine-tuning method that can effectively improve fairness in a wide range of debiasing methods.
  arXiv  Detail & Related papers  (2024-10-06T13:09:48Z)
• Spoken Stereoset: On Evaluating Social Bias Toward Speaker in Speech Large Language Models [50.40276881893513]
  This study introduces Spoken Stereoset, a dataset specifically designed to evaluate social biases in Speech Large Language Models (SLLMs) By examining how different models respond to speech from diverse demographic groups, we aim to identify these biases. The findings indicate that while most models show minimal bias, some still exhibit slightly stereotypical or anti-stereotypical tendencies.
  arXiv  Detail & Related papers  (2024-08-14T16:55:06Z)
• Identifying and Mitigating Social Bias Knowledge in Language Models [52.52955281662332]
  We propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases.<n>FAST surpasses state-of-the-art baselines with superior debiasing performance.<n>This highlights the potential of fine-grained debiasing strategies to achieve fairness in large language models.
  arXiv  Detail & Related papers  (2024-08-07T17:14:58Z)
• Dataset Scale and Societal Consistency Mediate Facial Impression Bias in Vision-Language AI [17.101569078791492]
  We study 43 CLIP vision-language models to determine whether they learn human-like facial impression biases. We show for the first time that the the degree to which a bias is shared across a society predicts the degree to which it is reflected in a CLIP model.
  arXiv  Detail & Related papers  (2024-08-04T08:26:58Z)
• Distilling Vision-Language Foundation Models: A Data-Free Approach via Prompt Diversification [49.41632476658246]
  We discuss the extension of DFKD to Vision-Language Foundation Models without access to the billion-level image-text datasets. The objective is to customize a student model for distribution-agnostic downstream tasks with given category concepts. We propose three novel Prompt Diversification methods to encourage image synthesis with diverse styles.
  arXiv  Detail & Related papers  (2024-07-21T13:26:30Z)
• FairCoT: Enhancing Fairness in Text-to-Image Generation via Chain of Thought Reasoning with Multimodal Large Language Models [14.632649933582648]
  We introduce FairCoT, a novel framework that enhances fairness in text to image models through Chain of Thought (CoT) reasoning.<n>We show that FairCoT significantly enhances fairness and diversity without sacrificing image quality or semantic fidelity.
  arXiv  Detail & Related papers  (2024-06-13T12:55:10Z)
• Detecting Bias in Large Language Models: Fine-tuned KcBERT [0.0]
  We define such harm as societal bias and assess ethnic, gender, and racial biases in a model fine-tuned with Korean comments. Our contribution lies in demonstrating that societal bias exists in Korean language models due to language-dependent characteristics.
  arXiv  Detail & Related papers  (2024-03-16T02:27:19Z)
• ChatGPT Based Data Augmentation for Improved Parameter-Efficient Debiasing of LLMs [65.9625653425636]
  Large Language models (LLMs) exhibit harmful social biases. This work introduces a novel approach utilizing ChatGPT to generate synthetic training data.
  arXiv  Detail & Related papers  (2024-02-19T01:28:48Z)
• Multilingual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May Not Help You [64.74707085021858]
  We show that multilingual models suffer from significant gender biases just as monolingual models do. We propose a novel benchmark, MAGBIG, intended to foster research on gender bias in multilingual models. Our results show that not only do models exhibit strong gender biases but they also behave differently across languages.
  arXiv  Detail & Related papers  (2024-01-29T12:02:28Z)
• Fair Text-to-Image Diffusion via Fair Mapping [32.02815667307623]
  We propose a flexible, model-agnostic, and lightweight approach that modifies a pre-trained text-to-image diffusion model. By effectively addressing the issue of implicit language bias, our method produces more fair and diverse image outputs.
  arXiv  Detail & Related papers  (2023-11-29T15:02:01Z)
• Social Bias Probing: Fairness Benchmarking for Language Models [38.180696489079985]
  This paper proposes a novel framework for probing language models for social biases by assessing disparate treatment. We curate SoFa, a large-scale benchmark designed to address the limitations of existing fairness collections. We show that biases within language models are more nuanced than acknowledged, indicating a broader scope of encoded biases than previously recognized.
  arXiv  Detail & Related papers  (2023-11-15T16:35:59Z)
• Fast Model Debias with Machine Unlearning [54.32026474971696]
  Deep neural networks might behave in a biased manner in many real-world scenarios. Existing debiasing methods suffer from high costs in bias labeling or model re-training. We propose a fast model debiasing framework (FMD) which offers an efficient approach to identify, evaluate and remove biases.
  arXiv  Detail & Related papers  (2023-10-19T08:10:57Z)
• Bias and Fairness in Large Language Models: A Survey [73.87651986156006]
  We present a comprehensive survey of bias evaluation and mitigation techniques for large language models (LLMs) We first consolidate, formalize, and expand notions of social bias and fairness in natural language processing. We then unify the literature by proposing three intuitive, two for bias evaluation, and one for mitigation.
  arXiv  Detail & Related papers  (2023-09-02T00:32:55Z)
• Steering Language Generation: Harnessing Contrastive Expert Guidance and Negative Prompting for Coherent and Diverse Synthetic Data Generation [0.0]
  Large Language Models (LLMs) hold immense potential to generate synthetic data of high quality and utility. We introduce contrastive expert guidance, where the difference between the logit distributions of fine-tuned and base language models is emphasised. We deem this dual-pronged approach to logit reshaping as STEER: Semantic Text Enhancement via Embedding Repositioning.
  arXiv  Detail & Related papers  (2023-08-15T08:49:14Z)
• Exposing Bias in Online Communities through Large-Scale Language Models [3.04585143845864]
  This work uses the flaw of bias in language models to explore the biases of six different online communities. The bias of the resulting models is evaluated by prompting the models with different demographics and comparing the sentiment and toxicity values of these generations. This work not only affirms how easily bias is absorbed from training data but also presents a scalable method to identify and compare the bias of different datasets or communities.
  arXiv  Detail & Related papers  (2023-06-04T08:09:26Z)
• UniDiff: Advancing Vision-Language Models with Generative and Discriminative Learning [86.91893533388628]
  This paper presents UniDiff, a unified multi-modal model that integrates image-text contrastive learning (ITC), text-conditioned image synthesis learning (IS), and reciprocal semantic consistency modeling (RSC) UniDiff demonstrates versatility in both multi-modal understanding and generative tasks.
  arXiv  Detail & Related papers  (2023-06-01T15:39:38Z)
• Debiasing Vision-Language Models via Biased Prompts [79.04467131711775]
  We propose a general approach for debiasing vision-language foundation models by projecting out biased directions in the text embedding. We show that debiasing only the text embedding with a calibrated projection matrix suffices to yield robust classifiers and fair generative models.
  arXiv  Detail & Related papers  (2023-01-31T20:09:33Z)
• Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models [103.61066310897928]
  Recent text-to-image generative models have demonstrated an unparalleled ability to generate diverse and creative imagery guided by a target text prompt. While revolutionary, current state-of-the-art diffusion models may still fail in generating images that fully convey the semantics in the given text prompt. We analyze the publicly available Stable Diffusion model and assess the existence of catastrophic neglect, where the model fails to generate one or more of the subjects from the input prompt. We introduce the concept of Generative Semantic Nursing (GSN), where we seek to intervene in the generative process on the fly during inference time to improve the faithfulness
  arXiv  Detail & Related papers  (2023-01-31T18:10:38Z)
• An Analysis of Social Biases Present in BERT Variants Across Multiple Languages [0.0]
  We investigate the bias present in monolingual BERT models across a diverse set of languages. We propose a template-based method to measure any kind of bias, based on sentence pseudo-likelihood. We conclude that current methods of probing for bias are highly language-dependent.
  arXiv  Detail & Related papers  (2022-11-25T23:38:08Z)
• D-BIAS: A Causality-Based Human-in-the-Loop System for Tackling Algorithmic Bias [57.87117733071416]
  We propose D-BIAS, a visual interactive tool that embodies human-in-the-loop AI approach for auditing and mitigating social biases. A user can detect the presence of bias against a group by identifying unfair causal relationships in the causal network. For each interaction, say weakening/deleting a biased causal edge, the system uses a novel method to simulate a new (debiased) dataset.
  arXiv  Detail & Related papers  (2022-08-10T03:41:48Z)
• Auto-regressive Image Synthesis with Integrated Quantization [55.51231796778219]
  This paper presents a versatile framework for conditional image generation. It incorporates the inductive bias of CNNs and powerful sequence modeling of auto-regression. Our method achieves superior diverse image generation performance as compared with the state-of-the-art.
  arXiv  Detail & Related papers  (2022-07-21T22:19:17Z)
• The Birth of Bias: A case study on the evolution of gender bias in an English language model [1.6344851071810076]
  We use a relatively small language model, using the LSTM architecture trained on an English Wikipedia corpus. We find that the representation of gender is dynamic and identify different phases during training. We show that gender information is represented increasingly locally in the input embeddings of the model.
  arXiv  Detail & Related papers  (2022-07-21T00:59:04Z)
• Does Data Repair Lead to Fair Models? Curating Contextually Fair Data To Reduce Model Bias [10.639605996067534]
  Contextual information is a valuable cue for Deep Neural Networks (DNNs) to learn better representations and improve accuracy. In COCO, many object categories have a much higher co-occurrence with men compared to women, which can bias a DNN's prediction in favor of men. We introduce a data repair algorithm using the coefficient of variation, which can curate fair and contextually balanced data for a protected class.
  arXiv  Detail & Related papers  (2021-10-20T06:00:03Z)
• Towards Understanding and Mitigating Social Biases in Language Models [107.82654101403264]
  Large-scale pretrained language models (LMs) can be potentially dangerous in manifesting undesirable representational biases. We propose steps towards mitigating social biases during text generation. Our empirical results and human evaluation demonstrate effectiveness in mitigating bias while retaining crucial contextual information.
  arXiv  Detail & Related papers  (2021-06-24T17:52:43Z)
• Towards Robustifying NLI Models Against Lexical Dataset Biases [94.79704960296108]
  This paper explores both data-level and model-level debiasing methods to robustify models against lexical dataset biases. First, we debias the dataset through data augmentation and enhancement, but show that the model bias cannot be fully removed via this method. The second approach employs a bag-of-words sub-model to capture the features that are likely to exploit the bias and prevents the original model from learning these biased features.
  arXiv  Detail & Related papers  (2020-05-10T17:56:10Z)

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.