VLMs show bias favoring text context over visual training data

A new research paper explores how vision-language models (VLMs) manage conflicting information between their training data and provided context. The study found that VLMs exhibit asymmetric biases, favoring text-based context information but relying on parametric (training) data for visual entities. This is attributed to the longer processing time for visual information, which hinders the suppression of existing knowledge. While chain-of-thought reasoning did not resolve this, increasing the amount of visual context did show an effect, highlighting challenges in achieving consistent behavior in multimodal and retrieval-augmented models. AI

IMPACT Highlights potential inconsistencies in multimodal AI systems, suggesting challenges for reliable information processing in complex applications.

RANK_REASON The cluster contains a single academic paper detailing research findings on AI model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• Athulith Paraselli
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• ScienceCast

• paper
• safety

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