Interpreting Latent CoT Reasoning as Dynamical Systems

Jul 14, 2026

Researchers model latent chain-of-thought (CoT) reasoning as dynamical systems, showing that CODI behaves as a stable attractor while COCONUT exhibits unstable, expanding dynamics. The study uses quantitative measures such as Lyapunov sensitivity and qualitative projections to demonstrate that latent CoT reasoning follows structured, non-random trajectories. SIM-CoT supervision is found to tighten both behaviors without altering their underlying dynamics.

Why it matters: This framework advances the interpretability of latent reasoning methods and provides actionable insights for improving their performance.

Full story at: arXiv AI/ML ↗