Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

debiaSAE: Benchmarking and Mitigating Vision-Language Model Bias

Kuleen Sasse, Shan Chen, Jackson Pond, Danielle Bitterman, John Osborne

TL;DR

This work systematically evaluates demographic bias in Vision-Language Models across five models and six datasets, revealing that portrait-based benchmarks like UTKFace and CelebA are effective for bias detection while scene-based benchmarks such as VLStereoSet and PATA often fail due to prompts that reveal answers without visual input. It also shows that VisoGender is too easy to challenge and that adversarial prompts expose hidden biases, underscoring the need for more robust benchmarks. To address these issues, the authors introduce a Sparse Autoencoder based debiasing method that steers latent representations and improves fairness by 5–15 points in adversarial and non-adversarial settings. The study demonstrates that bias evaluation in VLMs is fragile and current benchmarks can misrepresent fairness, while SAE-based debiasing provides a scalable, controllable path toward more equitable multimodal systems with practical implications for deploying VLMs in sensitive domains.

Abstract

As Vision Language Models (VLMs) gain widespread use, their fairness remains under-explored. In this paper, we analyze demographic biases across five models and six datasets. We find that portrait datasets like UTKFace and CelebA are the best tools for bias detection, finding gaps in performance and fairness for both LLaVa and CLIP models. Scene-based datasets like PATA and VLStereoSet fail to be useful benchmarks for bias due to their text prompts allowing the model to guess the answer without a picture. As for pronoun-based datasets like VisoGender, we receive mixed signals as only some subsets of the data are useful in providing insights. To alleviate these two problems, we introduce a more rigorous evaluation dataset and a debiasing method based on Sparse Autoencoders to help reduce bias in models. We find that our data set generates more meaningful errors than the previous data sets. Furthermore, our debiasing method improves fairness, gaining 5-15 points in performance over the baseline. This study displays the problems with the current benchmarks for measuring demographic bias in Vision Language Models and introduces both a more effective dataset for measuring bias and a novel and interpretable debiasing method based on Sparse Autoencoders.

debiaSAE: Benchmarking and Mitigating Vision-Language Model Bias

TL;DR

This work systematically evaluates demographic bias in Vision-Language Models across five models and six datasets, revealing that portrait-based benchmarks like UTKFace and CelebA are effective for bias detection while scene-based benchmarks such as VLStereoSet and PATA often fail due to prompts that reveal answers without visual input. It also shows that VisoGender is too easy to challenge and that adversarial prompts expose hidden biases, underscoring the need for more robust benchmarks. To address these issues, the authors introduce a Sparse Autoencoder based debiasing method that steers latent representations and improves fairness by 5–15 points in adversarial and non-adversarial settings. The study demonstrates that bias evaluation in VLMs is fragile and current benchmarks can misrepresent fairness, while SAE-based debiasing provides a scalable, controllable path toward more equitable multimodal systems with practical implications for deploying VLMs in sensitive domains.

Abstract

As Vision Language Models (VLMs) gain widespread use, their fairness remains under-explored. In this paper, we analyze demographic biases across five models and six datasets. We find that portrait datasets like UTKFace and CelebA are the best tools for bias detection, finding gaps in performance and fairness for both LLaVa and CLIP models. Scene-based datasets like PATA and VLStereoSet fail to be useful benchmarks for bias due to their text prompts allowing the model to guess the answer without a picture. As for pronoun-based datasets like VisoGender, we receive mixed signals as only some subsets of the data are useful in providing insights. To alleviate these two problems, we introduce a more rigorous evaluation dataset and a debiasing method based on Sparse Autoencoders to help reduce bias in models. We find that our data set generates more meaningful errors than the previous data sets. Furthermore, our debiasing method improves fairness, gaining 5-15 points in performance over the baseline. This study displays the problems with the current benchmarks for measuring demographic bias in Vision Language Models and introduces both a more effective dataset for measuring bias and a novel and interpretable debiasing method based on Sparse Autoencoders.

Paper Structure

This paper contains 54 sections, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Related Work
  4. Bias in LLMs
  5. Bias in VLMs
  6. Sparse Autoencoders: Steering and Transfer
  7. Methods
  8. Datasets
  9. Models
  10. Prompts and Prediction
  11. Dataset Effectiveness
  12. Adversarial Robustness
  13. Debiasing Method
  14. Metrics
  15. Bias Dataset Effectiveness Metrics
  16. ...and 39 more sections

Figures (5)

  • Figure 1: VLMs Demographic Bias Benchmark Taxonomy
  • Figure 2: Diagram demonstrating our process of creating the adversarial version of the VisoGender Dataset
  • Figure 3: Performance of models on all the portrait datasets compared to their fairness (DPR demographic parity ratio, higher the better.). The dashed diagonal line represents the dividing line between trading off performance for fairness or vice versa. Arrows indicate changes in performance and fairness from CLIP to LLaVa models when the differences for both are $>$ 10%. Note: PaliGemma did not use these CLIP based encoders.
  • Figure 4: Comparison between "Text Only" and "With Image" of average VLBS for VLStereoSet. Grey bar represents baseline CLIP model.
  • Figure 5: Average Demographic Parity Ratio comparison between "Text Only" and "With Image" for PATA (left). Macro-F1 comparison between "Text Only" and "With Image" for PATA (right). The grey bar represents the baseline CLIP Large 336 model, which is the backbone frozen vision encoder for all LLaVa that we have tested here.