marcusala233-ai/Thermodynamic-Phase-Transitions-in-Large-Language-Models-A-Spin-Glass-Perspective-on-Hallucination: Thermodynamic-Phase-Transitions-in-Large-Language-Models-A-Spin-Glass-Perspective-on-Hallucination

Large Language Models (LLMs) exhibit a sudden degradation in factual accuracy known as "hallucination." While often treated as a statistical error, we propose a thermodynamic framework where hallucination corresponds to a disorder-induced phase transition. Modeling the neural weights as frozen disorder in a Sherrington-Kirkpatrick spin glass, we demonstrate that the "temperature" hyperparameter controls a transition from a retrieval phase (ferromagnetic/glassy state) to a confabulation phase (paramagnetic state). We identify a critical temperature $T_c$ defined by the spectral radius of the weight matrix, offering a quantifiable metric for the onset of hallucination.