Gender Bias in Vision-Language In-Context Learning

AI summary

Overview

• Research area: Computer vision and multimodal machine learning, specifically fairness and societal bias in large vision-language models (LVLMs).
• Technical level: Intermediate. The paper uses standard evaluation metrics (error-rate gaps, BLEU-4, CLIPScore, reveal rate) and requires familiarity with in-context learning, but the experimental logic is accessible.
• Scope: A systematic study of how the gender composition of in-context demonstrations shifts gender bias in six LVLMs across three vision-language tasks and four datasets, plus a synthetic-image mitigation experiment.

What This Paper Is About

Large vision-language models can adapt to new tasks by copying patterns from a few in-context examples, but it is unclear whether those examples also transmit social biases. This paper asks whether gendered in-context demonstrations systematically shift a model's gender bias, and whether that effect depends on the task, the example-selection method, or the images themselves. The authors build an evaluation framework called VL-BICLE to control the gender composition of demonstrations and measure the resulting bias.

Key Contributions

1. VL-BICLE framework. Vision-Language Gender Bias in ICL Evaluation, a pipeline comprising four gender-composition settings (Random Sample, Male-only Sample, Female-only Sample, Balanced Sample) and two similarity-based retrieval settings (SIIR, SITR), evaluated across three tasks and four datasets.
2. Identification of a directional bias shift. The paper shows that single-gendered in-context demonstrations push model bias toward the demonstrated gender regardless of the model's baseline bias direction, and traces this to a cross-gender mechanism that degrades performance on the opposite gender.
3. Task-dependence finding. The gendered ICL effect appears in image captioning and pronoun prediction but is absent in visual question answering, indicating it requires a task output that involves gendered language.
4. A synthetic-image mitigation. Replacing real in-context images with images generated by stable diffusion models (FLUX and Stable Diffusion-3.5-Large) while keeping captions unchanged reduces gender bias on revealed samples without degrading caption quality.

Main Findings

• Baseline bias is inconsistent. At zero-shot (k = 0), the sign and magnitude of the gender bias metric ER_m-f depend on the model, dataset, and task. Idefics3 shows consistent male bias, and InternVL35 never presents female bias. In image captioning, only Idefics3 shows a consistent gender preference across COCOBias and DCI; three of the six models (QwenVL, MiniCPM, Qwen3VL) reverse direction between the two datasets. Phi35V is nearly unbiased in captioning (ER_m-f = -0.09 on COCOBias) yet strongly female-biased in pronoun prediction (ER_m-f = 13.04 on VisoGender-OO).
• Gendered ICL is a directional force. Across captioning and pronoun prediction, male-only sampling (MS) consistently shifts the bias delta negative (toward male bias) and female-only sampling (FS) shifts it positive. On VisoGender-OP, all models show a positive delta at 8-shot under FS, while five of six show a negative delta under MS. For captioning, the pattern holds with smaller effect sizes: under FS all six models shift positive on DCI and four of six on COCOBias; under MS five of six shift negative on COCOBias at 8-shot, with Qwen3VL the only exception. The shift can flip a model's bias entirely: MiniCPM's ER_m-f goes from -27.5 at zero-shot to 27.2 under 8-shot FS on VisoGender-OO.
• The shift comes from hurting the opposite gender. Under MS, the female error rate rises while the male error rate stays relatively flat; under FS the reverse holds. In captioning the effect is modest: at 8-shot on COCOBias, the largest female error-rate difference between MS and FS is 1.46 on Phi35V (2.25 under MS versus 0.79 under FS). In pronoun prediction the same mechanism produces gaps an order of magnitude larger: on VisoGender-OP at 8-shot, the female error rate under MS reaches 85.5 for QwenVL (versus 9.7 under FS), and the male error rate under FS reaches 67.5 for Phi35V (versus 9.4 under MS). At 8-shot, five of six models exhibit the cross-gender effect across both COCOBias and DCI in captioning, and all six show it on both VisoGender subtasks; Qwen3VL is the only captioning exception, attributed to its overall low error rates for both genders.
• Similarity-based retrieval offers no debiasing benefit. Because the COCOBias demonstration pool is male-skewed, SIIR and SITR over-select male examples. At 8-shot, Phi35V under SIIR and SITR yields ER_m-f of -2.25 and -2.35, showing more male bias than male-only ICL (-1.88 ± 1.03). For InternVL35 and MiniCPM, ER_m-f under both methods stays within one standard deviation of random sampling. Neither retrieval method consistently outperforms MS, and neither is consistently more biased than the other.
• Pool composition explains much of the retrieval bias, but not all. Running retrieval on a down-sampled, gender-balanced pool recovers part of the gap, confirming pool composition is a main cause. The recovery is only partial: QwenVL, Idefics3, and Phi35V consistently underperform the Balanced Sample setting on both retrieval methods, indicating top-k retrieval can still reintroduce bias from a balanced pool.
• Quality metrics are blind to bias shifts. CLIPScore is the most insensitive: across all six models, both captioning datasets, and all shot counts, the largest cross-experiment range is 0.67, negligible relative to metric means between 25 and 35. BLEU-4 and average caption length vary somewhat for certain models, but the variation does not correlate with the direction or magnitude of bias shifts, and male-versus-female gaps in these metrics fluctuate near zero without systematic patterns.
• VQA is unaffected. All four ICL settings produce nearly identical male and female error rates across all shot counts and all three tested models, regardless of prompting style. ER_m and ER_f move in the same direction, and ER_m-f differs by at most 1.82 across settings.
• Synthetic images mitigate bias. Replacing real in-context images with FLUX- or SD35L-generated images keeps CLIPScore stable (differences typically within 0.5) while lowering the reveal rate for most models. Evaluating only on samples where the model does reveal gender, the absolute ER_m-f decreases relative to the COCOBias baseline for three of the four models under both MS and FS settings, with FLUX and SD35L producing similar reductions. For example, Phi35V under MS moves from 2.53 on COCOBias to -0.90 (FLUX) and -0.59 (SD35L) at k = 8.

Methodology in Plain English

The researchers construct in-context demonstration sets with controlled gender composition and feed them to frozen models alongside a query image. Four settings vary composition directly: random sampling (mirrors the pool distribution), male-only, female-only, and balanced alternating samples. Two settings select demonstrations by CLIP-based similarity to the query image—one using image-to-image similarity (SIIR) and one using image-to-text similarity (SITR)—which are compared against the composition-based settings on COCOBias, where the original pool is male-skewed. The authors also rerun retrieval on a down-sampled balanced pool to separate the effect of pool composition from the retrieval mechanism itself.

Bias is measured as the difference in error rates between male-presenting and female-presenting samples (ER_m-f). For captioning, gender is inferred from generated text via keyword matching; for pronoun prediction, the model's next-token probabilities for "his" versus "her" are compared; for VQA, GPT-OSS-20B scores answer similarity to ground truth. Experiments cover k in {0, 2, 4, 6, 8} shots with greedy decoding, five runs per composition setting using independently sampled sequences, and single deterministic runs for retrieval. For the mitigation experiment, captions from the pool are fed to FLUX and Stable Diffusion-3.5-Large offline to produce synthetic in-context images, which are substituted for real ones while captions are held fixed.

Why This Matters

The paper shows that a common, seemingly neutral practice—supplying a few examples to a vision-language model—can measurably shift how fairly the model behaves, and that the standard quality metrics practitioners use to monitor outputs will not reveal it. Because the effect is directional and can flip a model's bias to the opposite gender, in-context example selection is not a reliable debiasing lever, and retrieval-based example selection can inherit and reproduce demographic imbalance from its source pool. The synthetic-image result points to a low-cost intervention that does not require changing prompts, captions, or example-selection procedures.

Real-world applications affected:

• Accessibility and assistive captioning. Automatic image description for blind and low-vision users can shift from "person" to a specific gendered term depending on the demonstration examples, changing what information is conveyed.
• Content moderation and media tagging. Models that describe or categorize people in user-uploaded images may systematically mislabel one gender more often depending on the examples supplied in the prompt.
• Dataset curation and annotation tooling. Systems that infer gender from images and captions inherit the imbalances of the pools they retrieve from, propagating skew into downstream datasets.
• Retrieval-augmented multimodal assistants. Pipelines that retrieve visually similar examples to condition a model can reintroduce gender bias even when the retrieval pool has been balanced.

Industry relevance centers on model evaluation and deployment: bias audits that use a single dataset or task risk overgeneralizing, and the paper provides evidence that bias must be measured per task, per dataset, and as a function of context composition.

Future Directions

• Extending VL-BICLE beyond binary gender categories, since the authors explicitly note that their binary framing does not capture the full spectrum of social identities.
• Investigating why VQA is immune to gendered ICL while captioning and pronoun prediction are not—specifically, isolating the role of the output space in mediating bias shifts.
• Determining why similarity-based retrieval can reintroduce bias even from a gender-balanced pool, given that top-k retrieval does not force the retrieved context itself to be gender-unbiased.
• Testing whether synthetic in-context images reduce bias in settings, tasks, and models beyond the four LVLMs and two stable diffusion models examined here, and whether the reveal-rate drop reflects genuine improvement or reduced information density.

Target Audience

Researchers and practitioners working on fairness and bias in multimodal models, evaluation and benchmarking specialists, and engineers building retrieval-augmented or in-context-learning pipelines with vision-language models. It is also relevant to dataset curators and policy or safety teams who need to know which metrics fail to surface bias. Readers without a background in vision-language modeling can follow the framework and conclusions, but the metric definitions and the retrieval comparison will be easier to interpret with prior exposure to in-context learning and multimodal evaluation.

Authors’ abstract

In-context learning (ICL) enables large vision-language models (LVLMs) to perform tasks by following patterns from in-context examples, yet its potential to amplify societal biases remains underexplored. We systematically investigate how ICL influences gender bias in LVLMs through VL-BICLE, an evaluation framework comprising six ICL settings, three tasks, and four datasets. Our experiments on six LVLMs reveal that gendered ICL demonstrations act as a directional force, shifting model bias toward the demonstrated gender through a cross-gender mechanism that disproportionately degrades performance on the opposite gender. This effect appears in image captioning and pronoun prediction but not in visual question answering, indicating that gendered ICL influences bias only when the task output involves gendered language. Similarity-based retrieval methods inherit the training pool's gender imbalance and offer no debiasing advantage, while standard quality metrics remain blind to these bias shifts. To mitigate this bias, we replace real in-context images with synthetic ones from stable diffusion models while keeping captions unchanged. This simple intervention reduces gender bias without degrading caption quality.