This is the problem with using loaded language like "reason" and "interpret". The model is not interpreting anything. All that is being done is a multdimentional map lookup with statistics.
> also don't see how that idea would fit in with the o1 models, which explicitly have "reasoning" tokens.
An LLM on top of an LLM (i.e using context to generate inputs to an LLM) is just a fancier LLM.
To really understand all of this, all you need to do is look at how Transformer works, namely the attention block. There is no such thing as Query, Key, and Value in the sense of how they are implied to be used. The may as well be called A,B,C, as they are all learned in training, and can be freely interchanged in naming. All you do for inference is multiply the output vector by A,B,C to get 3 matrices, then multiply them together (technically with a scaling factor for 2 of them, but again, doesn't matter for which 2, and the scaling factor can be built into the matrix itself)
And because you can unroll matrix multiplication into a 2 layer neural network, that means that any LLM in its current form today can be represented as a set of linear layers. And we know that a set of linear layers is simply a function. And every function has a finite range for a finite domain. And the inability to expand that range given a finite domain means its not reasoning.
So we have to rely on hacks like temperature to make it appear like reasoning, when its really not even close.