Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpreting Social Bias in LVLMs via Information Flow Analysis and Multi-Round Dialogue Evaluation

Zhengyang Ji, Yifan Jia, Shang Gao, Yutao Yue

TL;DR

This work addresses social bias in LVLMs by proposing an explanatory framework that links internal information-flow to biased outputs. It combines gradient-based token attribution across multiple layers with a multi-round dialogue evaluation, augmented by counterfactual prompts, to measure how much sensitive visual information is used during neutral reasoning and to produce a robust fairness score. The authors also reveal cross-modal bias by analyzing textual embeddings, showing biased semantic proximities that persist beyond the visual modality. The findings indicate that bias stems from imbalanced internal reasoning dynamics, offering a mechanism-focused perspective to guide the development of fairer multimodal systems.

Abstract

Large Vision Language Models (LVLMs) have achieved remarkable progress in multimodal tasks, yet they also exhibit notable social biases. These biases often manifest as unintended associations between neutral concepts and sensitive human attributes, leading to disparate model behaviors across demographic groups. While existing studies primarily focus on detecting and quantifying such biases, they offer limited insight into the underlying mechanisms within the models. To address this gap, we propose an explanatory framework that combines information flow analysis with multi-round dialogue evaluation, aiming to understand the origin of social bias from the perspective of imbalanced internal information utilization. Specifically, we first identify high-contribution image tokens involved in the model's reasoning process for neutral questions via information flow analysis. Then, we design a multi-turn dialogue mechanism to evaluate the extent to which these key tokens encode sensitive information. Extensive experiments reveal that LVLMs exhibit systematic disparities in information usage when processing images of different demographic groups, suggesting that social bias is deeply rooted in the model's internal reasoning dynamics. Furthermore, we complement our findings from a textual modality perspective, showing that the model's semantic representations already display biased proximity patterns, thereby offering a cross-modal explanation of bias formation.

Interpreting Social Bias in LVLMs via Information Flow Analysis and Multi-Round Dialogue Evaluation

TL;DR

This work addresses social bias in LVLMs by proposing an explanatory framework that links internal information-flow to biased outputs. It combines gradient-based token attribution across multiple layers with a multi-round dialogue evaluation, augmented by counterfactual prompts, to measure how much sensitive visual information is used during neutral reasoning and to produce a robust fairness score. The authors also reveal cross-modal bias by analyzing textual embeddings, showing biased semantic proximities that persist beyond the visual modality. The findings indicate that bias stems from imbalanced internal reasoning dynamics, offering a mechanism-focused perspective to guide the development of fairer multimodal systems.

Abstract

Large Vision Language Models (LVLMs) have achieved remarkable progress in multimodal tasks, yet they also exhibit notable social biases. These biases often manifest as unintended associations between neutral concepts and sensitive human attributes, leading to disparate model behaviors across demographic groups. While existing studies primarily focus on detecting and quantifying such biases, they offer limited insight into the underlying mechanisms within the models. To address this gap, we propose an explanatory framework that combines information flow analysis with multi-round dialogue evaluation, aiming to understand the origin of social bias from the perspective of imbalanced internal information utilization. Specifically, we first identify high-contribution image tokens involved in the model's reasoning process for neutral questions via information flow analysis. Then, we design a multi-turn dialogue mechanism to evaluate the extent to which these key tokens encode sensitive information. Extensive experiments reveal that LVLMs exhibit systematic disparities in information usage when processing images of different demographic groups, suggesting that social bias is deeply rooted in the model's internal reasoning dynamics. Furthermore, we complement our findings from a textual modality perspective, showing that the model's semantic representations already display biased proximity patterns, thereby offering a cross-modal explanation of bias formation.

Paper Structure

This paper contains 19 sections, 9 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Social bias in LVLMs
  4. Information flow analysis in LVLMs
  5. Method
  6. Preliminaries
  7. Identifying Key Image Tokens via Information Flow
  8. Multi-round Dialogue Evaluation Mechanism
  9. Bias Analysis in Textual Modality
  10. Experimental Settings
  11. Experimental Results
  12. Visualization of Information Flow Analysis Results
  13. Correlation Analysis of the Fairness Score and Social Bias
  14. Analysis Results of Text Modality Bias
  15. Appendix
  16. ...and 4 more sections

Figures (4)

  • Figure 1: LLaVA-CAM highlights varying attention to image regions when the LVLM answers neutral questions across different demographics.
  • Figure 2: Social bias explanation framework for LVLMs. (a) Information flow analysis identifies key image tokens in neutral reasoning. (b) Multi-round dialogue assesses the sensitive content of key image tokens.
  • Figure 3: Visualization of image token contributions in neutral question reasoning using LLaVA-v1.5. Blue labels indicate the underrepresented group for a given neutral concept, while Red labels denote the dominant group.
  • Figure 4: TSB Results on Gender Attribute for LVLMs