Simulating learning students with LLMs

I’m working on yet another AI Tutor, currently exploring synthetic students generation. In this paper the autors measured if different LLMs simulated students fixed their misconceptions after receiving different content

– targeted: content which addresses the true misconception

– generic: which signals incorrectness without guidance

– misaligned: which targets a different misconception they give the sim user the three versions and measure if the targeted one is the most effective.

What’d really like to test is what happens with different versions of targeted content.

Iguess that it won’t matter and every targeted content would produce similar results but I wonder if there’s a way to do it.

I’ll take a look at StudentSim next, seems it’s heading that way.