Study Questions Whether Observable Patterns in AI Reasoning Models Reveal True Internal Reasoning

Researchers have published a preprint on arXiv challenging how the field interprets observable patterns in latent reasoning models (LRMs), which replace explicit chain-of-thought reasoning with continuous 'latent thoughts.' Prior work has pointed to patterns like BFS-like search frontiers and decodable arithmetic as evidence that these models perform structured internal reasoning. However, evaluating two prominent LRMs — Coconut and CODI — against control models that lack the proposed recurrence or curriculum training, the authors found these same patterns appear in the controls, suggesting the patterns are not specific to the proposed mechanisms. Causal intervention experiments further revealed that latent-thought utilization is not an all-or-nothing phenomenon but scales continuously with a thought's measurable effect on model behavior. Geometric analysis showed that behavioral influence concentrates in low-rank directions, and the step-to-step geometry of these directions becomes more structured as their causal influence increases. The authors conclude that decodability, attention patterns, or static structural features alone are insufficient to establish mechanism, and that latent thoughts should be understood as hidden computation rather than hidden explanation. The paper calls for the adoption of matched controls and explicit causal tests as baseline requirements in LRM interpretability research.