Maybe they are thinking about embedding a program generator and execution environment into their LLM inferencing loop in a tighter way. The model invents a program that guides the output in a specific/algorithmic way, tailored to the prompt.