New research identifies "visual lock-in" in vision-language models

Researchers have identified a phenomenon called "visual lock-in" in vision-language models, where outdated verbal descriptions can lead to incorrect decisions about current scenes. This occurs when changes in the model's internal representation, triggered by prior information, are concentrated along specific "Prior Directions." The study found that models exhibiting stronger lock-in have these changes organized in a compact, reusable pattern. Interventions that removed components aligned with these Prior Directions restored correct visual grounding, suggesting that the coherence of these patterns dictates how easily a model can update its understanding. AI

IMPACT Identifies a specific failure mode in vision-language models that could impact their reliability in dynamic environments.

RANK_REASON Academic paper detailing a new phenomenon in vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• arXivLabs
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• GUI Grounding
• Hugging Face
• Influence Flower
• Prior Directions
• ScienceCast
• vision-language model

• paper
• model release

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