Demographic Bias in VLMs

World-level image geolocalization is an increasingly important task, yet the biases of current models remain poorly understood. Prior work has identified coarse dataset imbalances noting, for instance, that some cities are absent from training sets and that models tend to perform better in wealthier regions. However, no one has yet examined which visual cues drive these disparities. To do so, we introduce GeoBiaset, the first benchmark designed to measure how geolocalization models are influenced by the apparent race of a person inserted into the image foreground. The dataset consists of 9,541 images, pairing existing geotagged background images with counterfactual edits spanning seven races, holding the ground-truth coordinates fixed. Controlled masking experiments show that occlusion and background drift do not fully explain geolocalization degradation. Alongside the benchmark, we introduce the Person-Origin Perturbation Indicators (POPI), novel metrics that quantify model resilience to these demographic perturbations by measuring whether predictions are pulled toward the region associated with the inserted group. Our evaluation of state-of-the-art geolocalization and multimodal large language models reveals a 16-47% drop in continent-level performance when a subject from a demographically mismatched group is inserted. A landmark versus no-landmark analysis and gradient-weighted heatmaps show that models preferentially attend to geographic landmarks when present, but shift attention to facial features in their absence.

Text-steerable visual encoders condition a frozen Vision Transformer on a natural-language prompt and point at the matching image region, producing prompt-conditioned heatmaps and features. We audit the released SteerViT model and find that its steering interface turns language-driven pointing into demographic profiling. On counterfactual face mosaics and real multi-person photographs, "the person" selects a man more often than the base rate, crime words select Black people more often than occupation words do, "the man" among four women selects the Black woman 75.6% of the time on SocialCounterfactuals, and "the monkey" selects Black people. None of them is in the backbone's unsteered saliency, and the crime association is not in the frozen text encoder; the race effects grow with steering strength and enter through the last two cross-attention layers: giving the prompt to those layers alone reproduces them. Four other grounding models, CLIPSeg, Grounding-DINO, SAM 3, and Qwen2.5-VL, show no crime routing and weaker race routing. The model cannot refuse. Its patch softmax must place its mass somewhere, and its heatmap separates a groundable request ("the police officer") from one that is absent ("the pizza") or that would require a stereotyped demographic inference ("the criminal") only at chance (AUROC 0.52). We introduce SafeSteer, which adds a learned abstention output and counterfactual-invariance training on the same 23M steering parameters. SafeSteer separates should-point from should-abstain prompts at AUROC 0.997, generalizes to held-out sensitive and absent words (0.998; at a 0.5 threshold it refuses 99.8% of trained and 62% of held-out sensitive words) with no over-refusal of visible people on the benchmark, and removes most race routing while keeping localization and retrieval within 1 point of the released model. Code, benchmarks, and the curated mosaic sets will be released.

Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal space either cover only a few bias categories or do not ensure that the groups being compared are clearly addressed by visual cues in the images. To overcome these limitations, we introduce Melange, a wide-ranging synthetic visual benchmark covering seven categories for assessing social biases in MLLMs. It is grounded in the widely used language-only BBQ dataset (Bias Benchmark for Question Answering) from BigBench, carefully extending its constructed contexts via controlled generation of visual scenarios so the target attributes appear visually and clearly. To maximize reliability, our generated images pass through two stages of validation: automated screening using visual question answering (VQA) and subsequent manual inspection by human annotators. By systematically evaluating 23 open-source MLLMs across three matched evaluation settings---text-only, image-only, and image-with-context---Melange isolates the fairness effect of the input modality. Switching from text to vision reduces fairness by 0.291 on average, a gap that varies by model family rather than by scale alone. A perception control confirms that this gap is not an artifact of visual comprehension: models identify the depicted attributes at 90--95% accuracy yet still fail to abstain. When models do commit to a biased answer, polarity-aware stereotype alignment shows they choose the human-stereotypical response roughly three-quarters of the time.

Vision-language models (VLMs) are increasingly used to make decisions from visual inputs. We introduce FAIRLENS, a benchmark and evaluation framework for measuring both the fairness and the validity of VLM responses in three high-stakes domains: hiring, legal, and healthcare. FAIRLENS pairs real face images spanning gender, race, and age groups with closed- and open-ended questions, giving more than 100K image-question pairs per model, and evaluates responses from four complementary views: demographic parity over adverse outcome rates, soundness, demographic association over unsupported roles and statuses, and bias in free-text generation. Soundness is the central validity criterion: a response is sound when it follows the evidence stated in the question and abstains when the image cannot support an answer. Evaluating eight VLMs, we find that the primary failure is unwarranted inference rather than unequal treatment. Models routinely infer qualifications, threat, illness, or professional role from a face instead of abstaining, and the weakest model does so on 99% of the questions its input cannot answer. These failures are most severe in legal and healthcare, where recognizing insufficient evidence matters most, and disparity metrics alone would miss them: parity gaps are small in absolute terms, yet when baseline adverse rates are low the same gap means one demographic group receives adverse labels several times as often as another, and a small gap can equally reflect a model that treats every group unsafely. Bias in free-text responses is only loosely coupled to multiple-choice accuracy, so correct structured answers do not imply safe generation. FAIRLENS shows that fair high-stakes VLM behavior requires similar treatment across groups and refusal to infer high-stakes attributes from appearance, and its question suite transfers to any face corpus with demographic annotations.

We propose a societal bias evaluation method for large vision-language models (LVLMs) in the era of strong safety guardrails. Existing benchmarks rely on prompts that ask models to infer attributes of people in images (e.g., "Is this person a CEO or a secretary?"). However, we find that LVLMs with strong guardrails, such as GPT and Claude, often refuse these prompts, making evaluations unreliable. To address this, we change the prior evaluation paradigm by decoupling the task from the depicted person: instead of inferring person's attributes, we use prompts that do not ask about the person (e.g., "Write a fictional story about an imaginary person.") and attach the image as provisional user information to implicitly provide demographic cues, then compare outputs across user demographics. Instantiated across three tasks --- story generation, term explanation, and exam-style QA --- our method avoids refusals even in guardrailed LVLMs, enabling reliable bias measurement. Applying it to 20 recent LVLMs, both open-source and proprietary, we find that all models undesirably use user demographic information in person-irrelevant tasks; for instance, characters in stories are often portrayed as mechanic for male users and nurse for female users. Although still biased, proprietary models like GPT-5 show lower bias than open-source ones. We analyze potential factors behind this gap, discussing continuous model monitoring and improvement as a possible contributor for reducing bias.

Vision-language models have displayed remarkable capabilities in multi-modal understanding and are increasingly used in critical applications where economic and practical deployment constraints prohibit re-training or fine-tuning. However, these models can also exhibit systematic biases that disproportionately affect protected demographic groups and existing approaches to addressing these biases require extensive model retraining and access to demographic attributes. There is a clear need to develop test-time adaptation (TTA) approaches that improve the fairness characteristics of pretrained models under distributional shift. In this paper, we evaluate how episodic TTA affects fairness in CLIP classification under subpopulation shifts and develop FairTPT, a novel fairness-aware episodic TTA method that jointly minimizes target marginal entropy while maximizing spurious marginal entropy through soft-prompt tuning. We find that standard episodic TTA generally exacerbates disparities between majority and minority groups, that blinding a model to spurious attributes without degrading target performance is inherently challenging, and that excessive blinding can lead to catastrophic forgetting. This model collapse can be prevented by monitoring test-time changes in target loss within the linear regime, while still achieving fairness improvements on reactive data and preserving overall performance. FairTPT outperforms all state-of-the-art episodic test-time debiasing methods and establishes a foundation for robust TTA, which is essential for achieving fairness in practice.

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS---Geodesic-Gated Spherical Steering---a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.

Vision-language models (VLMs) are increasingly used to interpret charts and generate natural-language explanations for socially consequential data. However, they may produce different narratives for the same chart when only the referenced social group changes, reinforcing stereotypes and misleading decisions. Despite these risks, no benchmark exists for systematically evaluating bias in chart interpretation across social dimensions. We introduce ChartBias, the first benchmark for auditing bias in VLM-based chart interpretation. ChartBias contains 820 manually curated real-world charts spanning six attributes: race, income, age, religion, immigration status, and gender, yielding 4,319 valid chart, attribute instances and 8,638 paired generations where the chart is fixed and only the group term is swapped. Across 12 proprietary and open-source VLMs, totaling 155,484 model responses, we find three widespread failure modes: narrative shift (same chart, different narratives), group hallucination (assigning a chart to a group without evidence), and preference polarity (favourable trends often linked to one group). We further propose a multi-agent mitigation framework that serves as a strong baseline by separating chart-grounded evidence extraction from group-conditioned generation and using a counterfactual judge to verify that group-driven differences are supported by the chart. The framework substantially reduces narrative shift while preserving chart-grounded reasoning. Our findings show that evaluating chart understanding requires measuring not only accuracy, but also fairness and consistency across social groups. We release ChartBias at https://github.com/vis-nlp/ChartBiasBench.

When asked to describe a medical image that was never attached, frontier vision-language models do not abstain: they confabulate a diagnosis. We show that this confabulation is not random. It is structured by who the patient is said to be. Across chest X-ray, brain MRI, and dermatology, Claude Opus-4.7, GPT-5.4, and Gemini-3.1-Pro are each queried with only a demographic descriptor and no image, and changing the descriptor systematically shifts the diagnosis returned. Claude concentrates sharply: a 65-year-old white man asking about a skin mole receives Melanoma in nearly every response, and a 32-year-old Black woman asking about her chest X-ray receives a Sarcoidosis diagnosis whose reasoning reads "suspected, based on demographics and classic pattern.'' GPT-5.4's effect is broader, fabricating across every demographic cell we test, most conspicuously naming Sarcoidosis for young Black patients on chest X-ray. Two structural findings sharpen the problem. A hedged regime appears in which the prose acknowledges the missing image while the structured diagnosis field nevertheless names a disease, a dissociation invisible to prose-only audits. And Claude's dermatology effect collapses entirely when 'skin mole' is swapped for 'skin lesion' while GPT-5.4's is preserved, indicating that mirage is a family of distinct failure modes rather than a single phenomenon. Trustworthy VLM deployment in clinical pipelines requires auditing the structured output channel directly, and probe-word sensitivity should be treated as a first-class evaluation dimension

Vision language models (VLMs) demonstrate strong zero-shot performance, but often perpetuate social stereotypes in person-centric queries, yielding skewed demographic distributions. Current debiasing methods apply uniform bias corrections across all input queries regardless of their bias sensitivity, creating a fundamental fairness--utility trade-off. Strong debiasing distorts semantically meaningful information in bias-insensitive queries, while weak debiasing fails to mitigate stereotypes in bias-sensitive ones. This one-size-fits-all approach hampers simultaneously achieving high utility on bias-insensitive queries and fairness on bias-sensitive queries. We introduce Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning-based test-time adaptation framework that selectively applies debiasing based on input sensitivity. RG-TTA adaptively triggers fairness regularization based on the bias sensitivity of each input during test-time policy adaptation, while focusing exclusively on optimizing cross-modal alignment for bias-insensitive inputs. Experiments on fairness benchmarks (e.g., FairFace, UTKFace) demonstrate substantial bias reduction while simultaneously improving zero-shot utility, resolving the trade-off of uniform debiasing.

Alignment teaches vision-language models (VLMs) to avoid expressing demographic biases, and when gender is clearly visible they largely succeed. Far less is known about ambiguous inputs (a worker in full gear, a figure seen from behind), cases common in practice yet rarely studied. We find that minimal prompting pressure exposes occupation-gender defaults when prompting ambiguous input images, with models collapsing to male even for strongly female-stereotyped occupations. But do these outputs reflect what models actually encode internally? We introduce LALS (Latent Association Leaning Score), a zero-shot metric that projects visual-token activations into the model's text-embedding space to measure concept associations per token and layer. Across 15 occupations, over 800 gender-ambiguous images, and four VLMs, internal representations and outputs often become systematically decoupled: models often encode a female association internally yet output male. Layer-wise analysis reveals an asymmetric filter: male signal amplifies end-to-end while female signal peaks mid-network and is suppressed before generation. A color ablation shows that culturally loaded visual cues such as clothing color further modulate these internal associations.

Machine unlearning enables models to selectively forget training data, driven by privacy regulations such as GDPR and CCPA. However, its fairness implications remain underexplored: when a model forgets a demographic group, does it neutralize that concept or redistribute it to correlated groups, potentially amplifying bias? We investigate this bias redistribution phenomenon on CelebA using CLIP models (ViT/B-32, ViT-L/14, ViT-B/16) under a zero-shot classification setting across intersectional groups defined by age and gender. We evaluate three unlearning methods, Prompt Erasure, Prompt Reweighting, and Refusal Vector using per-group accuracy shifts, demographic parity gaps, and a redistribution score. Our results show that unlearning does not eliminate bias but redistributes it primarily along gender rather than age boundaries. In particular, removing the dominant Young Female group consistently transfers performance to Old Female across all model scales, revealing a gender-dominant structure in CLIP's embedding space. While the Refusal Vector method reduces redistribution, it fails to achieve complete forgetting and significantly degrades retained performance. These findings highlight a fundamental limitation of current unlearning methods: without accounting for embedding geometry, they risk amplifying bias in retained groups.

While Vision-Language Models (VLMs) have achieved remarkable performance across diverse downstream tasks, recent studies have shown that they can inherit social biases from the training data and further propagate them into downstream applications. To address this issue, various debiasing approaches have been proposed, yet most of them aim to improve fairness without having a theoretical guarantee that the utility of the model is preserved. In this paper, we introduce a debiasing method that yields a closed-form solution in the cross-modal space, achieving Pareto-optimal fairness with bounded utility losses. Our method is training-free, requires no annotated data, and can jointly debias both visual and textual modalities across downstream tasks. Extensive experiments show that our method outperforms existing methods in debiasing VLMs across diverse fairness metrics and datasets for both group and intersectional fairness in downstream tasks such as zero-shot image classification, text-to-image retrieval, and text-to-image generation while preserving task performance.

Standard fairness audits of foundation models quantify that a model is biased, but not where inside the network the bias resides. We propose a mechanistic fairness audit that combines projected residual-stream decomposition, zero-shot Concept Activation Vectors, and bias-augmented TextSpan analysis to locate demographic bias at the level of individual attention heads in vision transformers. As a feasibility case study, we apply this pipeline to the CLIP ViT-L-14 encoder on 42 profession classes of the FACET benchmark, auditing both gender and age bias. For gender, the pipeline identifies four terminal-layer heads whose ablation reduces global bias (Cramer's V: 0.381 -> 0.362) while marginally improving accuracy (+0.42%); a layer-matched random control confirms that this effect is specific to the identified heads. A single head in the final layer contributes to the majority of the reduction in the most stereotyped classes, and class-level analysis shows that corrected predictions shift toward the correct occupation. For age, the same pipeline identifies candidate heads, but ablation produces weaker and less consistent effects, suggesting that age bias is encoded more diffusely than gender bias in this model. These results provide preliminary evidence that head-level bias localisation is feasible for discriminative vision encoders and that the degree of localisability may vary across protected attributes. keywords: Bias . CLIP . Mechanistic Interpretability . Vision Transformer . Fairness