Designing AI Peers for Collaborative Mathematical Problem Solving with Middle School Students: A Participatory Design Study

Collaborative problem solving (CPS) is a fundamental practice in middle-school mathematics education; however, student groups frequently stall or struggle without ongoing teacher support. Recent work has explored how Generative AI tools can be designed to support one-on-one tutoring, but little is known about how AI can be designed as peer learning partners in collaborative learning contexts. We conducted a participatory design study with 24 middle school students, who first engaged in mathematics CPS tasks with AI peers in a technology probe, and then collaboratively designed their ideal AI peer. Our findings reveal that students envision an AI peer as competent in mathematics yet explicitly deferential, providing progressive scaffolds such as hints and checks under clear student control. Students preferred a tone of friendly expertise over exaggerated personas. We also discuss design recommendations and implications for AI peers in middle school mathematics CPS.

💡 Research Summary

This paper investigates how generative AI can be designed to act as a peer rather than a tutor in middle‑school mathematics collaborative problem solving (CPS). While CPS is a cornerstone of middle‑school math instruction, groups often stall or become dominated by a single voice when teacher support is intermittent. Existing AI tools focus on one‑to‑one tutoring, providing step‑by‑step hints or feedback, and there is scant research on AI agents that participate as equals within a student group. To fill this gap, the authors define an “AI peer” as an artificial agent introduced as a fellow student teammate that contributes ideas, asks questions, and shares responsibility for the task, emphasizing co‑participation over authority.

The study employed a child‑centered participatory design (PD) approach embedded in a five‑day summer camp (15 hours total) with 24 middle‑school students (grades 6‑8). Each day consisted of (1) a human‑only CPS activity, (2) a “technology probe” where students interacted with a prototype AI peer (LLM‑based chatbot offering hints, error detection, and concept refreshers), and (3) co‑design sessions in which students created personas, scenarios, and interface sketches for their ideal AI peer. Data were collected via observation notes, interview transcripts, design artifacts, and post‑camp surveys. Qualitative thematic coding and frequency analysis were applied to extract students’ perceptions, desired features, and control mechanisms.

Key findings:

1. Role perception – Students want the AI to be mathematically competent but explicitly deferential, acting as a “friendly expert” rather than an authority figure.
2. Scaffolding style – Preference for a progressive, “hint‑first” approach: error detection followed by targeted hints, concept reviews, and verification prompts before any direct answer is given.
3. Control & transparency – Participants demand explicit controls over when the AI speaks, how much it says, and what type of help it provides (e.g., “request help” button, selectable hint levels, time‑boxed AI turns).
4. Tone & persona – A polite, knowledgeable tone is favored; excessive persona embellishments (avatars, emotive language) are seen as distracting, while a minimal “friend‑like” identity is acceptable.
5. Domain‑specific scaffolds – Desired AI functions include (a) concise concept summaries, (b) “Did I get this step right?” checks, and (c) follow‑up practice problems that reinforce the current topic.

From these insights the authors propose five design recommendations: (1) implement a tiered hint system that encourages productive struggle, (2) provide student‑driven UI controls for AI turn‑taking, (3) maintain a courteous yet expert tone, (4) keep persona minimal but personable, and (5) embed math‑specific scaffolds such as concept refreshers, verification prompts, and extra practice. They argue that AI peers should act as “collaboration facilitators” complementing, not replacing, teacher orchestration, and that future work should integrate teacher‑AI coordination mechanisms.

Contributions include (1) a replicable PD‑based camp protocol for co‑designing AI peers with middle‑schoolers, (2) an empirical account of how adolescents evaluate AI along collaborative, social, and affective dimensions, highlighting the tension between reduced task load and equitable participation, and (3) concrete design guidelines for next‑generation AI agents that support group‑based mathematics learning. Limitations noted are the short, single‑site study and the lack of longitudinal classroom evaluation. The authors call for prototype development, classroom pilots, and rigorous measurement of learning and interaction outcomes to validate and extend their findings.

Overall, the paper demonstrates that involving youth directly in the design process yields nuanced, context‑specific specifications for AI peers, moving AI‑in‑education beyond tutor‑centric models toward tools that genuinely enhance collaborative reasoning in middle‑school mathematics.