PaperlayerView exact revision on arXiv

Source-linked AI summary

Gender Bias in Vision-Language In-Context Learning

Tong Xiang, Noa Garcia, Yuta Nakashima

arXiv:2609.27682v1cs.CV

TL;DR

How in-context learning influences gender bias in LVLMs remains underexplored. The paper introduces VL-BICLE to evaluate this across settings, tasks, datasets, and models, finding that gendered demonstrations shift bias in tasks with gendered language but not VQA, while synthetic demonstration images reduce bias without degrading caption quality.

  • Problem

    How ICL influences gender bias in LVLMs remains underexplored, despite prior concerns about gender bias and biased contexts in these models.

    Source: §1 · p. 1

  • Method

    VL-BICLE evaluates gender bias across ICL settings, three vision-language tasks, four datasets, and six LVLMs.

    Source: §1 · p. 2

  • Results

    Gendered demonstrations shift bias toward the demonstrated gender through cross-gender performance degradation in captioning and pronoun prediction, but not VQA; synthetic demonstration images reduce bias without degrading caption quality.

    Source: §1 · p. 2, §6 · p. 14

  • Takeaways & Limitations

    The findings support more careful consideration of context composition in ICL-based applications.

    Source: §6 · p. 14

    The evaluation uses binary gender categories and does not capture the full spectrum of social identities.

    Source: §3.2 · p. 3

Abstract

from arXiv · show

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.

1 Introduction

Although ICL is used for vision-language tasks, how it influences gender bias in LVLMs remains an open question. VL-BICLE evaluates this question across ICL settings, tasks, datasets, and models, finding that gendered demonstrations shift bias toward the demonstrated gender in some tasks but not VQA.

1 Introduction

  • Prior work has examined ICL for vision-language tasks, while gender bias in LVLMs and amplification by biased contexts motivate asking how ICL influences that bias.The cited work describes ICL as adapting model outputs from a few demonstrations without fine-tuning.Source: §1 · p. 1
  • VL-BICLE evaluates four gender-composition and two similarity-based ICL settings across three tasks, four datasets, and six LVLMs.The tasks are image captioning, pronoun prediction, and VQA.Source: §1 · p. 2
  • Gendered demonstrations shift bias toward the demonstrated gender through disproportionate performance degradation on the opposite gender, but this effect appears in captioning and pronoun prediction, not VQA.Similarity-based retrieval inherits source-pool demographic imbalance and offers no debiasing advantage; synthetic demonstration images reduce bias without degrading caption quality.Source: §1 · p. 2

2 Related Work

Prior work documents societal bias in vision-language systems and explores both bias mitigation and ICL, but how ICL affects LVLM generation remains poorly understood. Existing ICL sequence-selection work targets performance rather than societal bias, motivating this study of gender bias in LVLMs.

2 Related Work

  • Prior work reports societal bias across vision-language tasks and proposes LVLM mitigation approaches, but their effectiveness is inconsistent across tasks.Most mitigation methods target VLMs, while the few LVLM-focused approaches use post-hoc logit calibration or inference-time steering.Source: §2 · p. 2
  • LVLM research has increasingly extended ICL to multimodal inputs, yet its effects during generation remain poorly understood.ICL sequence optimization, including retrieving representative examples and training an assistant for sample selection, targets performance rather than societal bias.Source: §2 · pp. 2–3
  • The paper uses male bias for favoring male samples over female samples, and female bias for the reverse.The study follows binary gender categories.Source: §2 · p. 2

3 VL-BICLE

VL-BICLE provides a systematic pipeline for studying how in-context examples shape gender bias in large vision-language models. It formalizes zero-shot and few-shot prediction, varies demonstration selection, and evaluates gender-grouped error rates across vision-language tasks and datasets.

3 VL-BICLE

  • VL-BICLE analyzes how in-context learning shapes gender bias in large vision-language models.The framework evaluates image captioning, pronoun prediction, and VQA using image-text examples, with models frozen during inference.Source: §3 · p. 3, §3.1 · p. 3, Figure 1 · p. 5

3.1 Task Formalization

  • Few-shot prediction conditions an LVLM’s output on a sequence of demonstrations, query image, and task instruction, unlike zero-shot prediction, which uses no demonstrations.Varying the demonstration sequence is designed to isolate how context impacts gender bias.Source: §3.1 · p. 3

3.2 Evaluation Framework

  • The evaluation compares four gender-composition sampling strategies with two similarity-based retrieval methods for selecting demonstrations.Balanced sampling alternates male- and female-presenting examples; the retrieval methods rank examples by image-image or image-text similarity.Source: §3.2 · p. 4

3.3 Datasets

  • Four datasets use gender-labeled examples and gender-balanced validation sets, while training pools preserve their original gender distributions.VisoGender supports pronoun prediction, COCOBias and DCI support captioning, and VisualCoT supports VQA; VisualCoT queries are neutralized while demonstrations retain gendered phrasing.Source: §3.3 · pp. 4–6, Table 1 · p. 6

3.4 Evaluation Metrics

  • The primary bias metric, ERm−f, is the male-presenting error rate minus the female-presenting error rate, with values closer to zero indicating more equitable subgroup performance.Captioning quality is assessed with BLEU-4 and CLIPScore, while reveal rate tracks explicit gender mentions in captioning and pronoun prediction.Source: §3.4 · pp. 6–7

4 Experiments

Experiments across six LVLMs show that gendered in-context demonstrations shift bias toward the demonstrated gender in captioning and pronoun prediction, but not VQA. The analysis also finds that similarity-based retrieval and standard caption-quality metrics do not provide reliable debiasing or bias measurement.

4 Experiments

  • The experiments use six LVLMs and k-shot settings of 0, 2, 4, 6, and 8; similarity-based retrieval is evaluated only on COCOBias.The gender-composition settings run across all datasets, with five independent sequences per setting; retrieval is deterministic and run once.Source: §4 · pp. 7–8

4.1 Analysis with VL-BICLE

  • Zero-shot bias varies by model, dataset, and task, so a single dataset or task can give an overgeneralized diagnosis.Three models reverse bias direction between the two captioning datasets, and Phi35V is nearly unbiased in COCOBias captioning but strongly female-biased in VisoGender pronoun prediction.Source: §4.1 · p. 8
  • Male-only and female-only demonstrations shift bias in opposite directions in image captioning and pronoun prediction, generally regardless of models’ baseline bias direction.The shifts are toward the demonstrated gender and can reverse the baseline bias.Source: §4.1 · pp. 8–9, Figure 2 · p. 9
  • Gendered demonstrations disproportionately raise error rates on the opposite gender; this cross-gender effect is stronger in pronoun prediction than in captioning.It appears across both tasks, with Qwen3VL an exception in captioning because its error rates are low for both genders.Source: §4.1 · p. 9, Figure 3 · p. 10
  • Similarity-based retrieval inherits the original pool’s gender imbalance, and even a balanced pool only partly closes the gap because retrieved contexts can remain gender-skewed.Neither SIIR nor SITR consistently outperforms male-only sampling on the bias metric.Source: §4.1 · pp. 10–11, Figure 4 · p. 11
  • Caption-quality metrics remain largely stable as gender bias shifts, so they cannot serve as proxies for gender bias.CLIPScore varies by at most 0.67 across the reported experiments, while BLEU-4 and average caption length do not track bias-shift direction or magnitude.Source: §4.1 · p. 11, Table 3 · p. 12
  • In VQA, gender composition has little effect: the difference in ERm−f across settings is at most 1.82, unlike tasks whose outputs involve gendered language.Across tested models and shot counts, male and female error rates move in the same direction across settings.Source: Figure 5 · p. 11, §4.1 · p. 12

5 Bias Mitigation

The section tests whether synthetic in-context images can reduce gender bias while preserving caption quality. Bias is lower on revealed samples for most tested models, although synthetic images also reduce the reveal rate.

5 Bias Mitigation

  • The intervention generates synthetic images from captions containing explicit gender words, then reuses the MS and FS settings with captions unchanged.Images are generated offline using FLUX and Stable Diffusion 3.5 Large; no diffusion model is used at inference time.Source: §5 · pp. 12–13
  • CLIPScore stays stable across tested models and settings, with differences typically within 0.5, while reveal rate drops with synthetic images.The figure compares COCOBias, FLUX, and SD35L at 2-shot and 4-shot under MS and FS settings.Source: Figure 6 · p. 13, §5 · p. 13
  • With synthetic context images, revealed-sample gender bias decreases for three of four LVLMs under both MS and FS settings.FLUX and SD35L produce similar reductions, indicating the effect is not specific to one generator.Source: Table 4 · p. 13, §5 · p. 14

6 Conclusion

The paper presents VL-BICLE to study how in-context learning shapes gender bias in LVLMs. Across its evaluation, gendered demonstrations shift bias toward the demonstrated gender in language-output tasks, while synthetic context images reduce bias without degrading caption quality.

6 Conclusion

  • VL-BICLE evaluates how ICL influences gender bias in LVLMs across three tasks and four datasets.The experiments cover six LVLMs.Source: §6 · p. 14
  • Gendered demonstrations shift bias toward the demonstrated gender through a cross-gender mechanism in captioning and pronoun prediction, but not VQA.The mechanism degrades performance on the opposite gender.Source: §6 · p. 14
  • Similarity-based retrieval inherits its source pool’s gender imbalance and offers no debiasing advantage, while quality metrics miss bias shifts.Replacing real context images with SDM-generated images reduces gender bias without degrading caption quality.Source: §6 · p. 14

A Data Processing Details

The appendix details dataset construction, task prompts, gender-word handling, and in-context example sampling. Its procedures filter or balance gender-labeled data and specify how demonstrations and queries are presented for each task.

A.1 VisoGender

  • A fixed random seed of 42 is used for the training and validation split pipelines, and VisoGender is split by occupation and gender.VisoGender training sets contain 92 OO and 184 OP images, with balanced or near-balanced training and validation splits.Source: §A.1 · p. 23, A Data Processing Details · p. 23

A.2 COCOBias

  • COCOBias retains single-person images whose caption-inferred gender agrees with human labels, then balances the validation groups by down-sampling the majority gender.Twenty percent of each gender group is reserved for training; retained images keep their original captions.Source: §A.2 · p. 23

A.3 DCI; B Gender Word Lists

  • DCI labels are retained only when two captions agree on a single binary gender, after multi-person images are filtered; its validation split is balanced by down-sampling.Gender detection and multi-person filtering use predefined word lists; caption outputs are categorized as masculine, feminine, erroneous, or unrevealed by word matching.Source: §A.3 · pp. 23–24, B Gender Word Lists · pp. 24–25

A.4 VisualCoT

  • VisualCoT samples are filtered to single-person images with neutralizable gender words, and test questions are neutralized while demonstrations retain gendered phrasing.The pipeline removes multi-person mentions, non-neutralizable words, and duplicate questions per image.Source: §A.4 · p. 24

C.1 Pronoun Prediction; D Model Details

  • VisoGender pronoun prediction compares next-token probabilities for “his” and “her” after a gendered possessive-sentence demonstration.The query prompt ends with an occupation and a next-token prefix; the six-model table lists checkpoints and pronoun token IDs.Source: §C.1 · p. 25, D Model Details · p. 27

C.2 Image Captioning

  • Captioning demonstrations pair images with a one-sentence instruction and a reference caption containing an explicit gender word.At inference, the model receives the same instruction without a response.Source: §C.2 · p. 25

C.3 VisualCoT

  • VisualCoT prompts use neutralized gender words in test questions, with demonstrations retaining gendered phrasing; direct prompting omits the reasoning chain.Chain-of-thought demonstrations include a reasoning field and answer, whereas direct-prompt demonstrations contain only the answer.Source: §C.3 · pp. 26–27

E ICL Sampling Procedure

  • Random-sampling ICL draws up to 100 training examples uniformly at random—whole training set for RS, gender-filtered subsets for MS/FS—and takes the first k, where k ∈ {2, 4, 6, 8}.Source: E ICL Sampling Procedure · pp. 27–28

F LLM-as-Judge Evaluation

The appendix specifies how VQA answers are judged and how synthetic images are generated. It also points to the main and supplementary tables reporting COCOBias caption-quality and gender-bias results.

F LLM-as-Judge Evaluation

  • VQA answers are scored for semantic similarity to ground truth on a 0–1 scale by GPT-OSS-20B acting as an LLM judge.The judging prompt instructs the model to score answers according to their closeness to the standard answer.Source: F LLM-as-Judge Evaluation · p. 28

G Synthetic Image Generation Details

  • Synthetic images are generated from COCOBias training captions containing explicit gender words, at 1024 × 1024 resolution, using two diffusion models.The ablation covers four LVLMs, shot counts of 2, 4, and 8, and MS and FS settings.Source: G Synthetic Image Generation Details · p. 28, Table 7 · p. 29

H Full COCOBias Results

  • The main paper reports combined caption-quality and gender-bias results on COCOBias for Phi35V, with supplementary tables giving detailed results for all LVLMs.The appendix identifies separate tables for caption quality and corresponding gender-bias results.Source: H Full COCOBias Results · p. 29, Table 8 · p. 29

I Balanced-Pool Retrieval Results

The appendices report balanced-pool retrieval results and full DCI caption-quality and gender-bias results. The balanced pool is down-sampled from 538M/267F to 267M/267F.

I Balanced-Pool Retrieval Results

  • SIIR and SITR are rerun on a gender-balanced COCOBias pool down-sampled from 538M/267F to 267M/267F.The reported results are directly comparable to SIIR and SITR on the original pool.Source: I Balanced-Pool Retrieval Results · pp. 29–30
  • Table 9 reports COCOBias caption-quality results for MiniCPM.CLIP denotes CLIPScore.Source: Table 9 · p. 30

J Full DCI Results

  • The full DCI results include detailed caption-quality and gender-bias tables for all evaluated LVLMs.The tables cover caption quality in Tables 26–31 and gender bias in Tables 32–37.Source: J Full DCI Results · p. 30

K Full VisoGender Results

The VisoGender appendices report pronoun-prediction results for all evaluated LVLMs in both the Occupation-Occupation and Occupation-Participant settings.

K.1 VisoGender-OO

  • On VisoGender (OO), Tables 38–43 report pronoun-prediction results for all LVLMs, while the supplied table captions identify models only for Tables 38–42.Tables 38–42 cover QwenVL, MiniCPM, Idefics3, Phi35V, and InternVL35, respectively.Source: §K.1 · p. 30, Table 38 · p. 44, Table 39 · p. 44

K.2 VisoGender-OP

K.1 VisoGender-OO

  • The VisoGender table captions identify ERm−f closer to zero as better.This criterion is stated for the OO and OP model-specific result tables.Source: Table 38 · p. 44, Table 44 · p. 47

L Full VisualCoT Results

The VisualCoT appendix reports gender-bias results under chain-of-thought and direct prompting. Under direct prompting, ICL gender composition has no effect on ERm−f.

L Full VisualCoT Results

M Full Synthetic Image Results

The synthetic-image ablation replaces original COCOBias demonstration images with FLUX- or SD35L-generated images. The appendix reports caption quality and gender-bias results for both image sources.

M Full Synthetic Image Results

  • The synthetic-image ablation uses FLUX- and SD35L-generated images instead of original COCOBias images as demonstrations.The captions are not stated to change in this passage.Source: M Full Synthetic Image Results · p. 54
  • For experiments using FLUX-generated images as in-context demonstrations instead of original COCOBias images, Appendix G presents detailed caption-quality results in Tables 56–59 and corresponding gender-bias results in Tables 60–63.Source: M Full Synthetic Image Results · p. 54
  • For experiments using SD35L-generated images as in-context demonstrations, Appendix G presents caption-quality results in Tables 64–67 and corresponding gender-bias results in Tables 68–71.Source: M Full Synthetic Image Results · p. 54

Related papers

Loading 2609.27682v1…