Deciphering Functions of Neurons in Vision-Language Models

Vision-language models contain specialized neurons with distinct roles: The research identified three types of neurons in VLMs—visual neurons (processing images), text neurons (processing language), and multi-modal neurons (processing both)—revealing that these AI systems organize information in interpretable, specialized ways rather than as a "black box."

AI transparency can be systematically improved: The researchers developed an automated framework using GPT-4o to explain what individual neurons do, demonstrating that it's feasible to decode the internal functions of complex AI systems—a crucial step toward making AI systems more trustworthy and accountable.

Most neurons serve specialized functions: The statistical analysis of LLaVA (a representative VLM) showed that neurons tend to specialize in either visual or text processing rather than handling mixed information, suggesting these models have clearer, more interpretable internal structures than previously understood.

Verification methods for AI explanations are necessary: The team developed an "activation simulator" to test whether their explanations of visual neurons were reliable, highlighting that responsible AI requires not just interpreting models, but validating those interpretations are accurate.

Open-source model transparency benefits responsible AI: By studying openly available VLMs and publishing interpretability methods, this work contributes to the broader goal of making advanced AI systems more transparent and trustworthy across diverse applications.