Beyond Problem Solving: Large Language Models for Emotional and Reflective Support in Mathematics Learning

Synthesis: Rief et al. (2026) extend an Intelligent Tutoring system for algebra — "Math with Matt" — beyond purely cognitive Scaffolding by adding an LLM-driven socio-emotional layer that targets math anxiety. The mindful version pairs a Pedagogical Agent chat with context-sensitive emotional support and guided breathing with mindful feedback and hint messages (not merely evaluative). In a classroom study with 252 seventh graders (42 retained after disruptions), the tool reduced executive state-math anxiety and improved learning overall, but no significant differences emerged between the mindful and cognitive-only conditions. However, students in the mindful condition reached comparable learning with less time and fewer requested hints — higher learning efficiency and more balanced problem solving — and rated the agent as more supportive and caring. The study positions LLMs as a scalable, adaptive socio-emotional layer inside cognitive math tutoring, without requiring specialized instructor training.

The Cognitive-Only Bias of Intelligent Tutoring

Intelligent Tutoring Systems (ITSs) traditionally concentrate their adaptive support on the cognitive side of learning — hints, feedback, and task sequencing — while rarely attending to the learner's emotional state during problem solving. This is a real gap because math anxiety, a fear-and-tension response to numerical and Problem Solving situations, impairs working memory, lowers performance, and reduces engagement. State-math anxiety in particular (the context-dependent reaction that occurs while working) is negatively associated with performance. Prior affective interventions, such as instructor-led guided breathing at the start of class, reliably reduce anxiety but rely on trained, multi-session programs that are hard to embed in regular instruction or digital environments. The authors set out to test whether large language models could deliver such socio-emotional support automatically, in real time, and at scale.

"Math with Matt": Layering Mindfulness onto Cognitive Tutoring

The system, "Math with Matt," keeps the cognitive tutoring machinery (adaptive algebra hints and feedback) and adds a mindful intervention arm delivered through a pedagogical agent named Matt:

1. Mindful chat — an LLM-based chat offering context-sensitive emotional support, guided breathing exercises, and non-judgmental encouragement grounded in mindfulness principles.
2. Mindful feedback and hints — support messages that attend to the student's emotional experience rather than only evaluating the answer, intended to reduce anxiety and improve the learning experience.

Students interacted with one of two versions: the Mindful condition (with the mindfulness layer) or a Cognitive condition (cognitive hints and feedback only). The study was run across seven classrooms at an international school in Japan with 252 seventh graders (ages 12–13), randomly assigned within class; external disruptions (illness and a subway incident) reduced the analyzed sample to 42 students, balanced across conditions.

Findings: Efficiency, Experience, and Emotional Support

Both versions improved students' math learning from pre- to post-test and reduced executive state-math anxiety, with no significant difference between conditions on these primary outcomes. The mindful interventions nonetheless produced a distinct pattern of results:

1. Greater learning efficiency — mindful-condition students reached a similar level of algebra learning with less learning time and fewer requested hints than the cognitive group.
2. More balanced problem solving — their solution behavior was less help-dependent, consistent with more adaptive self-regulation during problem solving.
3. Stronger perceived care — students rated the single item "I feel supported by Matt" significantly higher in the mindful condition, indicating the agent felt more supportive and caring.

The authors note that limited English proficiency among students may have masked stronger between-condition effects, since learners could not always fully comprehend the mindful language in hints and chat.

What the study contributes

The study demonstrates the feasibility of integrating mindfulness into an ITS through LLM-based interaction: emotional scaffolding can be generated on demand, without trained instructors or fixed multi-session protocols. The pattern of equal learning with fewer hints and less time suggests the mindful layer may promote learning efficiency and self-regulated problem-solving behavior rather than merely improving how the tool "feels." Because LLM interactions can be localized and adapted in real time, they offer a scalable route to the kind of socio-emotional support that cognitive tutors have historically lacked — a direction with implications for mathematics education, learner Well-Being, and the design of classroom AI tutoring.

What this means for practice

• Instructors. Judge a socio-emotional layer by efficiency, not just by anxiety scores: both conditions improved algebra learning and reduced executive state-math anxiety (significant time effect, F(1, 15) = 5.422, p = .034, partial η² = 0.27), but mindful-condition students reached comparable learning with less time and fewer requested hints.
• Instructors. Track hint requests and time on task alongside pre/post test scores when you evaluate a tutoring change; those were the measures that separated the two conditions.
• Instructors. Do not run guided breathing as an in-class group ritual without structuring it. The authors observed that the classroom setting suppressed individual, self-paced mindfulness practice (students made fun of others) and recommend assigning such activities as homework.
• Designers. Fold emotional support into the existing hint and feedback loop rather than bolting on a separate chat: mindful-condition students rated "I feel supported by Matt" significantly higher, indicating that supportive wording in routine messages carries perceived care.
• Designers. Localize the language and check learners' proficiency before evaluating an LLM socio-emotional layer; here limited English proficiency likely masked the intervention's effects because the mindful language in hints and chat was not always understood.

Limitations

• Only 42 of the 252 participating seventh graders were analyzed (18 mindful, 24 cognitive), because illness and a major subway incident canceled sessions in five of the seven classes at one international school in Japan; the authors state the smaller-than-expected sample may not generalize across contexts.
• Students' limited English proficiency meant interface elements, hints, feedback, and chat messages were not always understood, and translation-tool use disrupted their workflow — a limitation the authors say likely masked the intervention's effects.
• Anxiety and perceived care are self-report measures: the Abbreviated Math Anxiety Scale (trait), the STAI-6 (state), and a single item, "I feel supported by Matt," for perceived care.
• The classroom context constrained the intervention itself: peer pressure limited engagement with individual, self-paced mindfulness practice, which the authors identify as a factor needing redesign (class-level randomization or homework delivery) in future studies.

Connected Concepts

• Intelligent Tutoring
• Math Education
• Pedagogical Agent
• Affective Computing
• Generative AI
• Large Language Models (LLMs)
• Self-Regulated Learning
• Well-Being
• K-12
• AI in Education

Connected Articles

• ProductiveMath: A Generative-AI-Powered App to Support Productive Failure Teaching — ProductiveMath: A Generative-AI-Powered App to Support Productive Failure Teaching
• Access is Not Enough: Human Support Improves Engagement with AI Tutoring — Access is Not Enough: Human Support Improves Engagement with AI Tutoring
• Adaptive Scaffolding for Cognitive Engagement in an Intelligent Tutoring System — Adaptive Scaffolding for Cognitive Engagement in an Intelligent Tutoring System
• From Prompting to Epistemic Proactivity: Temporal Trajectories of Student-AI Interaction in Mathematics Learning — From Prompting to Epistemic Proactivity in Mathematics Learning
• Curiosity as Linguistic Intervention: Using LLM Tutoring Dialogues to Influence Exploratory Learning Behavior — Curiosity as Linguistic Intervention with LLM Tutoring Dialogues

Citation

Rief, V., Hladký, M., Yoo, M., Heel, S., Sato, S., & Nagashima, T. (2026). Beyond Problem Solving: Large Language Models for Emotional and Reflective Support in Mathematics Learning. arXiv preprint arXiv:2609.02611.