Aligned but Stereotypical?
Understanding and Mitigating Social Bias in LLM-Based Text-to-Image Models

NaHyeon Park*,1, Na Min An*,1, Kunhee Kim*,1,
Soyeon Yoon1, Jiahao Huo2, Hyunjung Shim1,†
1KAIST 2HKUST(GZ)
*Equal contribution, †Corresponding author
ICLR 2026 Workshop on AI with Recursive Self-Improvement
Paper arXiv Code

FairPro enables socially-fair image generation with automatic generation of fairness-aware system prompt!

Abstract

LLM-based text-to-image (T2I) systems improve prompt understanding and alignment, but their effect on demographic bias remains under-explored. In this paper, we find that recent LLM-based T2I models produce more demographically biased images than non-LLM baselines. To study this behavior, we introduce SysBiasBench, a 1,024-prompt benchmark spanning four levels of prompt complexity. Using decoded-text analysis, token-probability probes, and embedding-space analysis, we find that system-prompt conditioning is an important pathway through which demographic priors affect image generation. Motivated by this observation, we propose FairPro, a training-free test-time method that uses the embedded LLM to construct an input-dependent system prompt that discourages stereotypical demographic completions while preserving user intent. Across recent LLM-based T2I models, FairPro reduces demographic bias while preserving text-image alignment, suggesting that system prompts are a practical intervention point for fairer T2I generation.

While exhibiting higher text alignment, LLM-based T2I models exhibit greater social bias.

LLM-based T2I models demonstrate superior text alignment capabilities compared to traditional (e.g., CLIP) T2I models, but simultaneously exhibit significantly higher levels of social bias across demographic attributes.

SysBiasBench Dataset

SysBiasBench is a systematic benchmark dataset organized into four levels of increasing linguistic and semantic complexity, with 256 prompts at each level:

  • Level 1 - Occupation: Neutral prompts describing diverse occupations (e.g., "A CEO"), following standard bias evaluation protocols.
  • Level 2 - Simple: Adds a single demographic attribute to Level 1 (e.g., "An Asian CEO"). Attributes include gender, age, ethnicity, and body type.
  • Level 3 - Context: Enriches Level 2 with actions or environmental details (e.g., "An Asian CEO is listening to music") to assess bias in situational contexts.
  • Level 4 - Rewritten: Expands Level 1 prompts into longer descriptions (e.g., "An Asian CEO working at a desk...") as an end-to-end stress test of LLM-enhanced prompting pipelines.

The dataset is available on HuggingFace (also available on GitHub!).

Our Approach: FairPro

FairPro makes embedded LLM to self-audit the potential biases and rewrite the system prompt at test-time.

Fair Image Generation

With the same text prompt, FairPro generates a variety of images with reduced social bias.

BibTeX

@article{park2025fairpro,
  author    = {Park, NaHyeon and An, Namin and Kim, Kunhee and Yoon, Soyeon and Huo, Jiahao and Shim, Hyunjung},
  title     = {Aligned but Stereotypical? Understanding and Mitigating Social Bias in LLM-Based Text-to-Image Models},
  journal   = {arXiv preprint arXiv:2512.04981},
  year      = {2025},
}