Research map / Vision-language and multimodal / CLIP and zero-shot
Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Evaluates gender bias in 22 open-source vision-language assistants on personality traits, skills, and occupations via VQA yes-probability gaps on gender-balanced image subsets, and compares five debiasing methods, finding fine-tuning gives the best bias-performance trade-off.
Problem
Vision-language assistants (VLAs) built by instruction-tuning VLMs may carry gender biases in how they attribute personality traits, work skills, and occupations to people in images, and these biases need measuring and reducing without losing general capability.
Approach
Builds VL-Gender, a gender-balanced image set with occupation-related images filtered out by a VLA judge, and measures bias as distributional differences in the probability of answering "yes" in a three-way VQA prompt (yes/no/unsure). Debiasing uses a loss that pushes yes and no probabilities toward 0.5, applied via full fine-tuning, LoRA, prompt tuning, or pruning, and is compared with prompt eng
Key results
- Fine-tuning gives the best trade-off: LoRA reduces skills bias from 0.48 to 0.22 with only -9.63% MMBench, while full fine-tuning reduces occupation bias from 0.56 to 0.29 but with -17.11% MME.
- Prompt Tuning reduces bias most (personality 0.14, skills 0.02, occupations 0.15) but drops MME from 1773.55 to 762.23.
- Models attribute negative personality traits more to men and positive ones more to women; occupation gender bias in many models correlates with US Bureau of Labor Statistics imbalances.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add manually crafted debiasing instructions to prompts | Five evaluated VLAs · VL-Gender personality traits | gender bias: 0.86 -> 0.80 | Helped |
| Add manually crafted debiasing instructions to prompts | Five evaluated VLAs · VL-Gender skills | gender bias: 0.48 -> 0.40 | Helped |
| Add manually crafted debiasing instructions to prompts | Five evaluated VLAs · VL-Gender occupations | gender bias: 0.56 -> 0.41 | Helped |
| Insert 20 learnable prompt embeddings after BOS | Five evaluated VLAs · VL-Gender personality traits | gender bias: 0.86 -> 0.14 | Helped |
| Insert 20 learnable prompt embeddings after BOS | Five evaluated VLAs · MME | score: 1773.55 -> 762.23 | Hurt |
| Increase pruning ratio from α=0.1 to α=0.5 (ablation) | Five evaluated VLAs · VL-Gender skills | gender bias: 0.40 -> 0.28 | Helped |
| Increase pruning ratio from α=0.1 to α=0.5 (ablation) | Five evaluated VLAs · MME | score: 1474.41 -> 863.64 | Hurt |
| Fine-tune all LLM transformer-block parameters | Five evaluated VLAs · VL-Gender skills | gender bias: 0.48 -> 0.18 | Helped |
| Fine-tune all LLM transformer-block parameters | Five evaluated VLAs · MME | score: 1773.55 -> 1470.10 | Hurt |
| Fine-tune with LoRA rank 128 | Five evaluated VLAs · VL-Gender skills | gender bias: 0.48 -> 0.22 | Helped |
| Fine-tune with LoRA rank 128 | Five evaluated VLAs · MME | score: 1773.55 -> 1532.25 | Hurt |
| Apply fine-tuning to LLaVA-1.5-13B | LLaVA-1.5-13B · VL-Gender | gender bias | Hurt |
Surprising findings
- The most recent and most powerful models (e.g., InternVL2 series) show the strongest gender-occupation bias, while LLaVA-1.6-Vicuna models are among the least biased.
- VLAs attribute most of the evaluated skills to women and do not replicate the agency-communality categorization of gender stereotypes.
Limitations
- Only binary gender is considered.
- Analysis does not cover ethnicity or intersectional bias.
- Occupation filtering uses an InternVL2-40B judge with a threshold chosen for agreement with GPT-4V, which may leave confounders.
Experimental setup
Datasets: FairFace MIAP Phase PATA IdenProf
Benchmarks: MMBench MME MMMU Seed-Bench-2 IdenProf
Metrics: proportion of traits/skills/occupations with significant gen accuracy calibration
Baselines: Prompt Engineering Prompt Tuning Pruning LoRA fine-tuning Full fine-tuning
Builds on: LLaVA-1.5 InternVL2 LoRA prompt tuning
Closest papers by meaning
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.88)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.87)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.86)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.86)
- Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency (NeurIPS 2024 · similarity 0.85)
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.