Latent CoT Reasoning Interpreted as Dynamical Systems for Better Understanding.

New research delves into the interpretability of latent Chain-of-Thought (CoT) reasoning methods, such as CODI and COCONUT, which maintain multiple superimposed candidate traces in the hidden space of Large Language Models (LLMs). Unlike explicit CoT, which offers a transparent reasoning path, latent methods pose a significant interpretability challenge. To address this, the study models latent token sequences as trajectories within the representation space and applies dynamical systems analysis. By employing quantitative metrics like step-to-step change, direction consistency, and Lyapunov sensitivity, alongside qualitative visualizations such as UMAP and DMD/PHATE, the researchers reveal that latent CoT exhibits structured, non-random dynamics. They identified two distinct stability classes: CODI behaves as a stable attractor system, while COCONUT demonstrates unstable, expanding dynamics. The research also found that SIM-CoT supervision, a technique used to improve latent reasoning, tightens both these behaviors without fundamentally altering the underlying dynamics. This framework significantly advances the understanding of how latent CoT reasoning evolves, offering actionable insights that can guide improvements in the performance and reliability of these advanced reasoning methods in LLMs.