A new research system uses an LLM to coach students through database design homework, but it is built to hold back answers rather than give them away.
The system plugs into an entity-relationship diagram editor and grounds its feedback in the student's actual diagram, the assignment rules, and rubrics written by instructors. Instead of one-shot grading, it runs a four-stage workflow: concept checks, guided application, low-detail feedback, and finally localized clarification. Each round of feedback is logged against a versioned snapshot of the diagram, so the system tracks how a student's work evolves after each nudge. In testing across three ERD tools and 383 feedback episodes, 71.1 percent of the changes students made afterward fully or partially matched the hidden diagnostic target the system was quietly tracking.
This matters because it is a concrete data point in the fight over what AI tutoring should withhold, not just what it should reveal. Most classroom chatbot tools either dump a full explanation or say nothing useful. Here, delayed disclosure sometimes worked as a feature: students avoided fixating on an inaccurate early hint, or picked up a concept without being handed the fix outright. That is a different design bet than the autocomplete-style assistance most ed-tech AI defaults to.
The catch is that students surveyed liked having more control over the process but also called it indirect and repetitive, and some mistakes still crept into revisions regardless of how the feedback was staged. Scaffolding is not the same as solving the underlying problem of students leaning on AI to think for them, it just changes how slowly the answer leaks out.