AI tutors that simulate how a specific student thinks are getting a way to remember without rereading everything from scratch.
Researchers built a framework called Learner2Skill that uses a large language model to simulate how a specific learner will respond to new problems. Most existing learner simulators feed the model that student's entire interaction history every time they need a new prediction, which gets slower and pricier as the history grows. Learner2Skill instead compresses what it has learned about a student's current understanding and recurring response patterns into a standalone file the researchers call a Simulation Skill, which updates as new interactions arrive and can transfer to a different large language model with only light recalibration rather than rebuilding the student profile from scratch. In the researchers' experiments, this approach reproduced fine-grained learner behavior more faithfully than replaying full histories, used fewer tokens overall, and the same Skill kept working when swapped between different LLM executors.
The real story here is portability, not accuracy. Ed-tech tools that lean on LLMs to model students are quietly locked into whichever model they started with; separating 'what this student knows' from 'which LLM is doing the reasoning' makes it cheaper to switch vendors or trim costs without rebuilding a simulator for every learner.
Worth noting: this comes from one paper testing the authors' own benchmarks, not an independent evaluation, and 'adapted to a new LLM' still means some recalibration work, not a free swap.