Mercury Functional Regions Across 38 Language Models

Key Points

• The aim is to identify and define recurring functional regions within various visual language models.
• Analyzed 38 visual language-model towers for recurring functional regions.
• Utilized metrics like occupancy, transition sharpness, and cross-model stability.
• Defined regions including embed/detokenization, surface, structural, semantic, and output zones.
• Establishes a measurable anatomy of models based on identified functional regions.
• Demonstrates cross-model stability in the defined regions.
• Highlights boundary uncertainty as a key characteristic of model behavior.