I guess one could just watch the token output stream and replace solvable expressions with calculated results, continuing with the replaced result. Kind of like

I guess one could just watch the token output stream and replace solvable expressions with calculated results, continuing with the replaced result. Kind of like the LLM producing a template, to be filled on the fly. Some models now do something like this for enforcing longer reasoning processes - they replace </think> with „Wait“.