MathBuddy: Affective Math Tutoring

Synthesis: MathBuddy: Affective Math Tutoring — EMNLP 2025 Demo. An emotionally aware Large Language Models (LLMs)-powered mathematics tutor that dynamically models student emotions from both conversational text and facial expressions, aggregating Multimodal AI affective signals to shape pedagogically appropriate LLM responses. Maps detected affective states to relevant pedagogical strategies, achieving a +23-point win rate advantage and +3-point improvement on the DAMR (Dimensions of Affective Math Response) score at the overall level. Evaluated across eight pedagogical dimensions in user studies, MathBuddy demonstrates that affective awareness substantially improves the quality of AI math tutoring.

Key Findings

MathBuddy addresses a critical gap in AI Tutoring: most AI math tutors respond to the content of student answers but ignore the emotional state of the learner. This is problematic because mathematics learning is deeply affective — frustration, confusion, anxiety, and confidence all shape how students engage with mathematical problems and whether they persist through difficulty.

Multimodal emotion modeling. MathBuddy captures student affect through two complementary channels: (1) conversational text analysis, which detects emotional signals in the linguistic content of student responses (e.g., expressions of frustration, confusion, or confidence), and (2) facial expression analysis, which captures moment-to-moment emotional reactions through computer vision. These modalities are aggregated to produce a unified affective state assessment that is more robust than either modality alone — a design principle consistent with the broader Multimodal Dialogue in STEM Education literature.

Affect-to-pedagogy mapping. The system maps detected affective states to specific Learning Design pedagogical strategies. For example, detected frustration may trigger scaffolded hint sequences or encouraging messages; detected confidence may prompt the tutor to increase problem difficulty or reduce Scaffolding; detected confusion may trigger a shift from procedural guidance to conceptual explanation. This dynamic mapping connects MathBuddy to the Affective Tutoring paradigm and the Learner Modeling and Adaptive Instruction tradition in Intelligent Tutoring.

LLM-powered response generation. At its core, MathBuddy uses an LLM prompted with the aggregated affective state, the mathematical context, and the tutoring history to generate emotionally appropriate pedagogical responses. The LLM is not merely an answer engine but an affective-pedagogical reasoning system that must balance mathematical correctness with emotional attunement.

Rigorous evaluation. The system was evaluated along eight pedagogical dimensions — including scaffolding quality, emotional attunement, mathematical accuracy, and learner engagement — through controlled user studies. The +23-point win rate and +3-point DAMR improvement establish clear evidence that affective awareness adds measurable value beyond content-only tutoring approaches. This connects to the engagement metrics literature and the growing recognition that Engagement Intensity as a Learner-Modeling Signal for Adaptive AI Ethics Instruction is essential for effective AI tutoring.

What this means for practice

• Instructors. Turn affect awareness on and act on it deliberately: the emotion-aware condition produced longer lasting positive emotional states and better-rated interactions, so treat the detected state as a cue to change strategy rather than as decoration.
• Designers. Aggregate text and facial signals into a single state before choosing a pedagogical strategy, and accept the recall cost the authors measured for the Neutral class when the modalities disagree.
• Designers. Plan for a third channel — spoken audio, handwritten notes, or biometrics — in technical subjects, where the authors found emotional expression in text too subtle for text plus face alone to be informative enough.
• Instructors. Log the inferred state across the session rather than reacting turn by turn: longer durations of positive facial emotion under the affect-aware condition are a persistence signal worth monitoring during hard problems.

Limitations

• The user study ran with 30 participants aged 15–55, recruited almost entirely from the authors' own university, so the sample includes no children — a gap the authors flag because younger learners' idea of ideal tutoring may differ substantially.
• Each participant completed only two tutoring sessions of roughly 10 minutes, and learning was measured with a 6-question multiple-choice test, so any learning effect is measured at a very coarse grain.
• Multimodal emotion aggregation reached just 60% accuracy against participants' own gold annotations, with Neutral recall of 0.15 and Positive recall of 0.41 — the system's read of affect is the weakest link in the pipeline.
• The authors state the field lacks a golden standard for tutoring quality, and their qualitative ratings rely on users who may rate user experience rather than educational prowess.

Citation

Kar, D., Böss, L., Braca, D., Dennerlein, S. M., Hubig, N. C., Wintersberger, P., et al. (2025). MathBuddy: Affective Math Tutoring.