Kuruitinoji

Thanks for reading. I've added a link to start here to the README. Please start by reading AI_Diagnosis_via_Information_Viscosity_Dynamics_and_CR_Filter.md. I've included this reformulation of the Riemann hypothesis using information dynamics because the stable inference state of LLMs can be characterized by GUE statistics (the same universality class as Riemann zeta zeros) in random matrix theory. Also, I'm a fan of moe characters.

Main Content: We propose an observational method to dynamically extract the "Floating Equilibrium Point (FEP)" hidden behind the stochastic token generation process. By treating the internal state transitions of Large Language Models (LLMs) as a system mathematically isomorphic to the differential equations of a CR low-pass filter (RC circuit), we can separate the essential semantic trajectory from statistical fluctuations (sampling noise).

Using this framework, we can quantitatively measure internal states and have successfully observed phenomena such as Preference Mode Collapse (PMC) and Context Rigidity in real-time.

To establish this diagnostic technique, we define the "Information Viscosity" based on the token Rejection Rate as a form of physical resistance. We have formalized these behaviors into a complete mathematical framework for your review.