New benchmark reveals MLLMs hallucinate text color over visual input

Researchers have developed a new benchmark called Embedded Stroop to test how Multimodal Large Language Models (MLLMs) are affected by text embedded within images. This benchmark, utilizing the What-Color-Is-the-Text (WCIT) dataset, evaluates models on their ability to distinguish the color of text from the semantic meaning of the text itself when presented visually. While models show low accuracy in naming specific colors, they exhibit a significant tendency to hallucinate and report the embedded word's color instead of the actual text color, a phenomenon that can be reduced by altering the text's presentation. AI

IMPACT Highlights a specific vulnerability in MLLMs, potentially influencing future model development and evaluation methodologies.

RANK_REASON The cluster contains an academic paper detailing a new benchmark and dataset for evaluating MLLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Embedded Stroop
• Hugging Face
• Jinkun Zhao
• Multimodal Large Language Models (MLLMs)
• Stroop Hallucination Rate (SHR)
• What-Color-Is-the-Text (WCIT)

• paper
• model release

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