cross-llm

This paper investigates whether different LLMs share common inference patterns when predicting the same token, using interaction-based explanations. Results show that advanced LLMs exhibit consistent interaction patterns, suggesting implicit optimization toward shared inference mechanisms.