This thing that people are calling “reasoning” is more like rendering to me really, or multi pass rendering. We’re just refining the render, there’s no reasoning involved.
Which is related to multistage/ hierarchical/coarse-to-fine optimization, which is a pretty good way to find the global optimum in many problem domains.