Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models

The paper presents a causal-geometric analysis of latent reasoning models (LRMs), specifically evaluating Coconut and CODI, revealing that observable patterns in latent states do not necessarily indicate internal reasoning mechanisms. The study demonstrates that these patterns also appear in control models lacking the proposed recurrence or curriculum, suggesting that latent thought utilization is graded rather than binary. This finding emphasizes the need for causal interventions and matched controls in LRM interpretability, as traditional metrics like decodability and attention do not adequately establish the underlying mechanisms.