Diagnosing why a student's math reasoning goes wrong usually requires asking the right follow-up questions and applying consistent criteria—something that is hard to sustain across a full classroom. The new system STEPwise, described in a recent arXiv preprint, addresses this by using large language models to generate questions and rubrics tailored to each student's work.

The tool is designed to support teachers rather than replace them: it produces individualized probing prompts and a rubric that helps interpret student responses. This lets teachers see where reasoning breaks down without having to design every diagnostic step themselves.

The paper argues that problem-solving is central to mathematics education, but that current classroom practice lacks the tools to probe each student's thinking consistently. STEPwise appears to be an attempt to close that gap, though the abstract does not yet provide details on evaluation or classroom results.