Vision-Language Models Show Text Bias in Stroop Task Experiments

Researchers have adapted the psychological Stroop paradigm to investigate how vision-language models (VLMs) handle images with conflicting features. When presented with images where text and color disagree, VLMs consistently prioritize the text feature over the color feature in their responses. Further analysis of the models' embedding spaces revealed that textual cues are represented more strongly and distinctly than color cues, which explains this bias and predicts the ease of intervening on text-related features. AI

IMPACT Highlights how feature saliency in VLMs can shape downstream application performance, suggesting a need for deeper understanding of representational biases.

RANK_REASON Academic paper detailing a novel experimental methodology applied to AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Hugging Face
• Stroop paradigm
• vision-language model

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →