Latent Chain-of-Thought Fails via Gradient Attenuation and Representational Drift
June 19, 2026
An information-theoretic analysis of latent CoT identifies two failure modes: gradient attenuation along the optimization path and semantic drift in hidden states. Decomposing process supervision into trajectory signals and generative space reconstruction outperforms rigid geometric compression for stabilizing latent reasoning.
HOW THIS AFFECTS YOU
●
researcherThe dual-collapse framing and the trajectory-vs-space supervision decomposition give concrete diagnostic handles for improving latent reasoning training in smaller models.