RUME 28
2026
Alexandria, Virginia
HaLLMos: Constitutional AI for Feedback on Mathematical Proofs
Page: 1168
Learning to construct and critique proofs is central to advanced mathematics but challenging for students and. This paper reports on an NSF-funded project to build HaLLMos, an AI model that helps students learn to write mathematical proofs. HaLLMos uses a constitutional approach along with expert-guided feedback to generate responses to students’ proof efforts. Guided by Schoenfeld’s Goals–Orientations–Resources framework and perspectives on proof as discourse, we are investigating the tool’s alignment with constitutional principles, as well as student and faculty perespectives on HaLLMos’ value. Preliminary data from interviews and system logs suggest that HaLLMos supports students’ independence while also preparing them to participate in mathematical conversations. Findings suggest HaLLMos can expand resources for students and faculty while raising new questions about how AI generated feedback can support students’ mathematical interactions and ultimately their sense of belonging in mathematics.