Interpreting Latent CoT Reasoning as Dynamical Systems

Recent latent reasoning methods such as CODI and COCONUT face a fundamental interpretability problem since they carry multiple superimposed candidate traces in hidden space at each step, obscuring how reasoning evolves, while explicit-CoT follows a single transparent reasoning trace at each step. Existing mechanistic methods show compression, shortcuts, and superposition without in-depth analysis, hence failing to explain how reasoning evolves across latent steps. To address this gap, we model the sequence of latent tokens as a trajectory in representation space and apply dynamical-system analysis to define the reasoning evolution. Using both quantitative (step-to-step change, direction consistency, Lyapunov sensitivity, etc.) and qualitative (UMAP and DMD/PHATE projections), we show that latent CoT exhibits structured, non-random dynamics with two distinct stability classes: CODI behaves as a stable attractor while COCONUT behaves as an unstable expanding system. Sim-CoT supervision tightens both behaviors handling latent instability in these methods without changing the underlying dynamics. This framework advances the interpretability of latent CoT reasoning dynamics and introduces actionable findings to catalyze further research into improving latent reasoning performance. All code, data, and other artifacts will be publicly released upon acceptance.