Mercury Cross-Model Highways: Navigating Shared Functional Addresses Across Language Models

Key Points

• This paper aims to define and inspect cross-model functional-address hypotheses in language models.
• Defines cross-model roads among opened LLM towers
• Utilizes region alignment and neighbor overlap for analysis
• Applies prompt co-activation, confidence scoring, and falsification tests
• Establishes a framework for inspecting functional addresses across models
• Introduces metrics for evaluating neighborhood overlap and alignment
• Demonstrates potential applications in optimizing language model interactions