An Information-Geometric Framework for Stability Analysis of Large Language Models under Entropic Stress

ArXi:2604.24076v1 Announce Type: new As large language models (LLMs) are increasingly deployed in high-stakes and operational settings, evaluation strategies based solely on aggregate accuracy are often insucient to characterize system reliability. This study proposes a thermodynamic inspired modeling framework for analyzing the stability of LLM outputs under conditions of uncertainty and perturbation. The framework