When LLMs Help -- and Hurt -- Teaching Assistants in Proof-Based Courses

Teaching assistants (TAs) are essential to grading and feedback provision in proof-based courses, yet these tasks are time-intensive and difficult to scale. Although Large Language Models (LLMs) have been studied for grading and feedback, their effectiveness in proof-based courses is still unknown. Before designing LLM-based systems for this context, a necessary prerequisite is to understand whether LLMs can meaningfully assist TAs with grading and feedback. As such, we present a multi-part case study functioning as a technology probe in an undergraduate proof-based course. We compare rubric-based grading decisions made by an LLM and TAs with varying levels of expertise and examine TAs' perceptions of feedback generated by an LLM. We find substantial disagreement between LLMs and TAs on grading decisions but that LLM-generated feedback can still be useful to TAs for submissions with major errors. We conclude by discussing design implications for human-AI grading and feedback systems in proof-based courses.