Trained models should be able to use formal tools (for instance a logical solver, a computer?).

Good. That said, I wonder if those models are still LLMs.