Research
State your model
Don't ask it to explain; explain it yourself, wrong, and let it correct you. A stated model turns vague answers into targeted checks.
The fastest way to learn from a machine is not asking it to explain; it is explaining to it, wrong, on purpose if need be, and letting it correct you. A stated model, in your own terms, however crude, converts the machine from lecturer to grader: instead of another fluent overview you get a targeted verdict on the exact place your understanding bends. Wrongness stated is progress; vagueness absorbed is not.
The same move rescues explanations that have stopped converging. When the answers circle, restating what you currently believe, in one ugly concrete analogy, forces the correction to attach to something. Crude analogies from your home domain are not embarrassing; they are instruments. The machine will meet you at your model and repair it in place.
How to practice it
Explain it back, badly.
After any explanation, restate it in your own terms and ask what's wrong with your version. The errors located are the lesson.
Anchor in your home domain.
Translate the new field into the one you know. The translation's failures map exactly where the new field surprises you.
Break the circle with a model.
When answers loop, stop asking and start asserting. A stated belief gives the correction something to grip.
Keep score on your model.
Track which parts of your stated model survived correction. What survived is what you actually know.