Coconut and CODI latent reasoning patterns not causal, study finds

Latent reasoning models (LRMs) like Coconut and CODI replace explicit chain-of-thought with continuous hidden states. Recent work treated observable patterns in these states—such as BFS-like frontiers or decodable arithmetic—as evidence of actual reasoning mechanisms. But a new causal-geometric analysis by Aswal, Palmeira Ferraz, Zhou, and Peyrard (arXiv:2606.12689) challenges that assumption. They evaluated two LRMs against carefully matched controls that lacked the recurrence or curriculum training the models rely on. Strikingly, the same patterns appeared in the controls, suggesting the patterns are not unique to reasoning. Causal interventions further revealed that utilization of latent thoughts is graded: a thought’s influence scales with its causal effect on behavior, and geometric analysis showed this effect concentrates in low-rank directions whose structure grows more organized as behavioral influence increases.

The paper delivers a clear warning: observable patterns are not explanations. Decodability, attention, or static structure alone cannot establish that a model is actually reasoning internally. Instead, interpretability for LRMs must rely on matched control experiments and causal interventions. This work directly impacts how researchers validate internal reasoning in hidden-state models—if patterns can be mimicked by simpler architectures, then the field needs more rigorous benchmarks. For professionals building on LRMs, the takeaway is that current interpretability claims might be overblown; future work should incorporate causal tests to separate genuine reasoning from spurious correlations.