AI/ ai · research · interpretability · machine-learning

Inside the Black Box of Latent AI Reasoning

New research applies dynamical systems analysis to latent chain-of-thought models, revealing two distinct stability classes that could guide better AI design.

Researchers have a new way to watch AI think — even when the thinking is invisible.

Latent reasoning methods like CODI and COCONUT skip the step-by-step text that makes standard chain-of-thought models legible. Instead, they juggle multiple candidate reasoning paths simultaneously inside hidden states — faster, but nearly impossible to audit. A new paper applies dynamical systems analysis to these latent token sequences, treating each reasoning step as a point along a trajectory in representation space. Using tools like Lyapunov sensitivity measures, UMAP projections, and a technique called Dynamic Mode Decomposition, the researchers found that latent reasoning is not random noise — it has structure, and that structure differs sharply between models.

The finding matters because interpretability in AI has mostly meant reading outputs, not dissecting the mechanics underneath them. This framework offers a vocabulary for the latter: CODI behaves like a stable attractor, converging toward a solution, while COCONUT behaves like an unstable expanding system, spreading outward. That distinction has real design implications — knowing whether a model is pulling toward an answer or drifting away from one changes how you would tune or supervise it.

The researchers also tested SIM-CoT supervision, which tightened both behaviors without altering their fundamental character — a result suggesting the underlying dynamics are baked in, not easily patched. For a field that has spent years arguing over whether large models "actually reason," this kind of mechanical evidence is more useful than another benchmark score.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →