5.6 and its ilk include a reasoning step; the output of the LLM is fed back through evaluation steps to determine if it's correct before emitting it, and these minimze some common issues with more naive LLM solutions (most notably the "confidently wrong" issue where it is over-fit on some lie and emits it).

One could argue that this is just more layers of LLM, but at the end of that argument is the observation that a 2-layer neural network with a nonlinear transform between the layers is a universal function; yes, but nobody does that because it's unrealistic to do so.