navi-SAD: A New Tool for Probing the Reasoning Mechanism of Large Language Models from a Dynamical Systems Perspective

Introduction / Main Floor: navi-SAD: A New Tool for Probing the Reasoning Mechanism of Large Language Models from a Dynamical Systems Perspective

navi-SAD is a reasoning monitoring tool for large language models based on dynamical systems theory. It measures the cosine divergence between softmax and linear attention through parallel computation, and reconstructs the attractor of the model's internal state using delay coordinate embedding technology, providing a new analytical perspective for understanding the reasoning behavior of LLMs.