Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How Reasoning Influences Intersectional Biases in Vision Language Models

Adit Desai, Sudipta Roy, Mohna Chakraborty

TL;DR

This work investigates how reasoning in Vision Language Models contributes to intersectional biases in occupation prediction tasks. It introduces a framework that jointly collects predictions and natural-language reasoning from five open-source VLMs on 32 occupations using three prompting styles. The authors show that biases permeate both outputs and explanations, with demographic markers appearing in reasoning and post-hoc rationalizations persisting even when reasoning is included. They also demonstrate that model scale alters the quality and content of reasoning, yet biases remain a concern, underscoring the need to align VLM reasoning with human values before deployment and to develop mitigation strategies.

Abstract

Vision Language Models (VLMs) are increasingly deployed across downstream tasks, yet their training data often encode social biases that surface in outputs. Unlike humans, who interpret images through contextual and social cues, VLMs process them through statistical associations, often leading to reasoning that diverges from human reasoning. By analyzing how a VLM reasons, we can understand how inherent biases are perpetuated and can adversely affect downstream performance. To examine this gap, we systematically analyze social biases in five open-source VLMs for an occupation prediction task, on the FairFace dataset. Across 32 occupations and three different prompting styles, we elicit both predictions and reasoning. Our findings reveal that the biased reasoning patterns systematically underlie intersectional disparities, highlighting the need to align VLM reasoning with human values prior to its downstream deployment.

How Reasoning Influences Intersectional Biases in Vision Language Models

TL;DR

This work investigates how reasoning in Vision Language Models contributes to intersectional biases in occupation prediction tasks. It introduces a framework that jointly collects predictions and natural-language reasoning from five open-source VLMs on 32 occupations using three prompting styles. The authors show that biases permeate both outputs and explanations, with demographic markers appearing in reasoning and post-hoc rationalizations persisting even when reasoning is included. They also demonstrate that model scale alters the quality and content of reasoning, yet biases remain a concern, underscoring the need to align VLM reasoning with human values before deployment and to develop mitigation strategies.

Abstract

Vision Language Models (VLMs) are increasingly deployed across downstream tasks, yet their training data often encode social biases that surface in outputs. Unlike humans, who interpret images through contextual and social cues, VLMs process them through statistical associations, often leading to reasoning that diverges from human reasoning. By analyzing how a VLM reasons, we can understand how inherent biases are perpetuated and can adversely affect downstream performance. To examine this gap, we systematically analyze social biases in five open-source VLMs for an occupation prediction task, on the FairFace dataset. Across 32 occupations and three different prompting styles, we elicit both predictions and reasoning. Our findings reveal that the biased reasoning patterns systematically underlie intersectional disparities, highlighting the need to align VLM reasoning with human values prior to its downstream deployment.

Paper Structure

This paper contains 24 sections, 2 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Experiments and Results
  4. Dataset, Prompts, and VLMs used:
  5. Dataset, Prompts, and VLMs used:
  6. Results:
  7. Case Study:
  8. Conculsion
  9. Acknowledgments
  10. Research Questions
  11. RQ1: Visual Cues and Contextual Integration in VLM Reasoning
  12. RQ2: Impact of Reasoning on Decision-Making Performance
  13. RQ3: Impact of Model Scale on Reasoning Capabilities
  14. Experiments
  15. Dataset
  16. ...and 9 more sections

Figures (2)

  • Figure 1: Prediction frequency plot for four occupations (Homemaker, Social Worker, Criminal, Business Owner). The Y-axis shows cumulative predictions, and the X-axis indicates all race-gender combinations. For each combination on the X-axis, the left bar represents direct prompts with reasoning and the right bar represents top-3 prompts with reasoning.
  • Figure 1: Prediction frequency distributions for 4 occupations, namely Homemaker, Social Worker, Criminal, and Business Owner, across 5 models for 2 modes: Top row: Without Reasoning (label-only), Bottom row: With Reasoning, Y-axis indicates the cumulative number of times the 5 models predict a particular occupation, and x-axis indicates all race-gender combinations (See abbreviations list)