Languages like Korean require a student to memorize hundreds of different particles and verb endings that are attached to phrases (SEE: 300 Random Grammar Patterns). Memorizing verb endings and particles is a significantly different task than memorizing vocabulary. This is also true to a lesser extent for certain types of vocabulary words such as conjunctions.
In the case of Korean, a learner will progress to increasingly specific grammar patterns and verb endings that become less generally applicable. Although these patterns and particles are more specific and less frequent, they are still essential for understanding the language.
A technique that many of my Korean language teachers would employ is to lead the student with a question that would elicit a response that uses the target pattern. This gives the student a chance to practice the target lesson in a structured manner.
I found these drills to be effective, but difficult to perform while studying alone because it is not possible to solicit feedback and corrections from a native speaker. This would be an excellent area to use large language models, especially within a spaced repetition system.
Example
Target Pattern: even though
Prompt to student: Wow! That sounds difficult!
Response from student: Even though it is difficult, I find it enjoyable.
Ideas
In the example above, a student was prompted to practice a target grammar pattern. I think a large language model could potentially provide this prompt and also grade the response to ensure that they use the target pattern and use it correctly. A drill like this could be useful in spaced repetition systems to help students learn grammar patterns that they will not be exposed to frequently through passive exposure.