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

In three linesCausal analysis of latent reasoning models (Coconut, CODI): observable patterns (BFS-like frontiers, decodable arithmetic) are not evidence of reasoning mechanisms. Causal interventions show latent-thought utilization is graded, not binary, and concentrated in low-rank directions. Decodability alone cannot establish mechanism.

Summary generated by Claude — human-verified