MathAIde: Teacher-Mediated Math ITS

• MathAIde is a mobile intelligent tutoring system that uses augmented intelligence to analyze handwritten elementary math, combining AI suggestions with teacher judgment.
• It employs computer vision and pre-defined correction models to classify student work, enhancing fairness and efficiency in low-resource classroom settings.
• The teacher-centered design, validated through classroom studies, shows improved accuracy and rapid, actionable feedback for early math education.

Searching arXiv for the specified MathAIde paper and closely related work to ground the article in current literature. MathAIde is a mobile Intelligent Tutoring System (ITS) for mathematics that supports handwritten input and is designed primarily for elementary-school math teachers working with students on early math skills. Its defining design choice is that students solve exercises on paper while the teacher acts as a proxy through a smartphone app: the teacher prepares or selects an exercise list, students answer in their notebooks, the teacher photographs each student’s work, and the system uses computer vision and AI to analyze the solution, judge whether it is correct, and produce feedback, analytics, and recommendations. In the study centered on the app, the distinctive contribution is not full automation but an augmented-intelligence correction mechanism in which AI suggests an interpretation of handwritten mathematical work while the teacher retains final assessment authority and can override or refine the system’s judgment (Guerino et al., 31 Jul 2025).

1. Definition, scope, and educational rationale

MathAIde is situated at the intersection of Intelligent Tutoring Systems, Artificial Intelligence in Education, and accessibility-oriented educational technology. The paper frames ITSs as among the most effective forms of AIED, but argues that three persistent challenges remain central: teachers are classroom decision-makers yet are often not involved throughout system design; AI tools are limited and unreliable; and access to conventional educational technology is unequal. Within that framing, MathAIde is presented as a teacher-mediated ITS aligned with “AIED unplugged,” because students continue to work on paper while the teacher bridges handwritten work and the AI system through a mobile interface (Guerino et al., 31 Jul 2025).

The app’s educational role is therefore narrower and more specific than that of a general-purpose mathematical assistant. It is not described as a student-facing conversational tutor, a cloud computer algebra system, or an autonomous competition solver. Instead, it is a classroom workflow system for elementary mathematics in which exercise assignment, handwriting capture, answer recognition, and review are mediated by the teacher. This suggests that MathAIde belongs most clearly to the ITS tradition in which information extraction, reasoning about student work, explanation or feedback, and adaptation are coordinated inside an educational workflow rather than exposed as separate standalone tools (Vaerenbergh et al., 2021).

A common misunderstanding is to treat MathAIde as an automated grader that replaces teacher judgment. The paper explicitly rejects that interpretation. Its central concept is Augmented Intelligence (AuI), defined as an approach in which AI enhances human capability rather than replacing human judgment. The crucial distinction is “decision-making authority”: the system can suggest or automate part of the analysis, but a person remains responsible for the final assessment (Guerino et al., 31 Jul 2025).

2. System architecture and operational workflow

The operational workflow described for MathAIde is both pedagogical and technical. A teacher first creates a list of exercises for a class, drawing on a repository of ready-made mathematics items already in the app. Students solve them on paper. The teacher then takes a photo of each student’s notebook. MathAIde analyzes the image using computer vision and AI and reports whether each answer appears correct or incorrect. Elsewhere in the broader platform, it also supports personalized feedback, learning analytics, and exercise recommendations, although those broader features are not the main analytical focus of the paper (Guerino et al., 31 Jul 2025).

The system is explicitly designed for low-resource settings. The paper states that the app operates with minimal requirements, supports data collection and analysis offline, and uses internet connectivity strategically for report updates and AI-model improvement. It also emphasizes role-awareness, teacher mediation, and simple interfaces to accommodate varied levels of technological skill. A plausible implication is that the design aims to preserve pedagogical continuity in environments where one-to-one student devices or stable connectivity are not available, without abandoning advanced AIED capabilities.

The clearest implementation detail concerns interface flow rather than model internals. In the deployed version, the “Record Execution” screen lets the teacher photograph the student’s notebook and link answers to exercises. On the “Linking” screen, the teacher receives student-linked information. A “Feedback” button opens the “Basic Report” screen, where each answer is shown separately and marked right or wrong according to the AI. If review is needed, the teacher selects “Report error,” which opens a drawer where the correction can be made (Guerino et al., 31 Jul 2025).

The underlying AI stack is described only at a moderate level of granularity. MathAIde uses “multiple AI features -- such as Object Recognition, Handwritten Math Equation Recognition, and Knowledge Tracing.” The paper does not provide a detailed pipeline diagram, exact OCR models, neural architectures, image preprocessing steps, confidence measures, loss functions, or formal scoring rules. The most concrete technical constraints are performance and scope limits: the system had “an overall accuracy of around 70\% in detecting all characters in a given answer and around 80\% in detecting all but one character,” and it was limited to addition, subtraction, and multiplication equations with up to three digits (Guerino et al., 31 Jul 2025).

This limited technical specification is important conceptually. It distinguishes MathAIde from formal execution engines such as MathPartner, which assume that a problem has already been correctly formalized in Mathpar or a LaTeX-like syntax and then solve it using symbolic-numerical methods (Malaschonok et al., 2024). MathAIde’s core challenge is earlier in the pipeline: turning messy, handwritten, classroom-produced notebook images into educationally actionable judgments.

3. Augmented intelligence as the core design principle

The paper’s central contribution is the design of an intervention mechanism for cases where handwritten mathematics recognition fails. The motivating problem is concrete: one student solved a problem correctly, but the app failed to identify the carry mark and interpreted the answer as 144 instead of 44; in another example, it failed to recognize a crossed-out digit 5 and interpreted it as 1. In both cases, correct work was marked incorrect. These examples are used to justify why MathAIde should not operate as a fully autonomous grader (Guerino et al., 31 Jul 2025).

The authors explicitly argue that full automation was not reliable enough for handwritten elementary mathematics in authentic classrooms. Recognition errors could arise because of photo quality, lighting, handwriting variation, camera quality, placement of digits, carry marks, crossed-out digits, and the open technical difficulty of handwritten math expression recognition. Replacing teacher judgment under those conditions would, in the paper’s framing, be pedagogically risky and would undermine trust. The chosen AuI design therefore aims to make the ITS “fairer and more assertive,” with teacher and AI “working together.”

Two alternative correction models were prototyped. Option A used a digitized-calculation editing model: the app displayed the recognized calculation, and the teacher could click on numbers to edit them, after which the app would re-analyze the calculation. Option B used a pre-defined-options model: the app displayed whether the student got the answer right or wrong and then let the teacher either indicate that the student actually got it right when the AI said they were wrong, or change the type of student error when the AI had identified a different one (Guerino et al., 31 Jul 2025).

The eventual choice of Option B reflects a general design principle of constrained flexibility. The paper repeatedly emphasizes the trade-off between fidelity to the student’s original work and classroom efficiency. Full digit-level correction preserved precision and was seen as potentially useful for AI learning, but it required too many interactions. Pre-defined remediation alternatives were less expressive but faster and less error-prone in actual use. This suggests that, in classroom ITS design, human override mechanisms may need to optimize for intervention cost rather than maximal representational richness.

That design stance is consonant with broader hybrid human-AI findings in mathematics support systems. In a separate study of hybrid human-AI tutoring in low-income middle-school settings, the central value of the human layer was not replacement of the AI layer’s instructional logic, but targeted intervention where software alone could not sustain engagement, accountability, or productive struggle (Thomas et al., 2023). MathAIde applies an analogous principle to assessment: the AI automates initial recognition and classification, but the human layer resolves ambiguity and preserves fairness.

4. Teacher-centered design methodology

Methodologically, the MathAIde paper is built around a mixed user-centered process with four studies: brainstorming with users, high-fidelity prototyping, A/B testing, and a classroom case study. The authors present this as a teacher-centered design approach in which teachers are involved in all phases rather than being consulted only at the end (Guerino et al., 31 Jul 2025).

The first study involved 14 elementary school teachers in a two-hour brainstorming session organized around the question, “How to correct a student’s answer that was misidentified by MathAIde?” The session produced six “valid ideas”: direct editing of misidentified numbers; fixed editing options such as “carry over was not detected” or “cut number was not detected”; declaring “reservations” in advance, such as telling the system that a problem requires carrying; imposing structured squares on answer sheets; using multiple-choice responses instead of free handwritten work; and using colors to signify digit roles or correction needs. The authors did not adopt all of these proposals, and they explicitly evaluated each in terms of pedagogical suitability, UX alignment, and technical feasibility.

That filtering process is significant. Some suggestions appeared technically helpful but pedagogically unacceptable. Answer-sheet squares might improve recognition, but were judged restrictive because they constrain children’s writing. Multiple choice was rejected because selecting the right option does not show correct calculation and is not appropriate for early grades in the same way. The resulting design process was therefore not simple participatory accumulation of features; it was iterative negotiation between classroom practice and machine-recognition constraints.

The second study, conducted with six internal project-team members, translated brainstormed ideas into implementable alternatives. The third study, an A/B test with three teachers who had not participated in brainstorming, used a fixed scenario in which the exercise was “Larissa had 2 flowers and received 3 more from her mother. How many flowers did Larissa end up with?”, the student answer was $2 + 3 = 4$, and the AI misidentified it as $2 + 3 = 7$. The timing results were decisive: Task 1 with Prototype A took on average 2m15s; Task 2 with Prototype B took 54s on average; and Task 3 with Prototype B took 41s. The paper states that “the time to use the solution was a determining factor in the proposal’s success” (Guerino et al., 31 Jul 2025).

Qualitative findings from that phase further shaped the final implementation. Teachers appreciated Prototype A because it was faithful to the student’s response and potentially useful for AI learning, but worried that it would consume too much classroom time. Prototype B was preferred because correction was easier and more directed, and because the options could generate “more reliable” data after passing through “2 filters (AI and teacher).” The main wording problem was that “Report Error” sounded like reporting a software bug rather than reviewing an AI judgment. On the basis of both timing and interview data, Prototype B was selected.

This teacher-centered design logic resonates with broader research on context-responsive AI in mathematics education. Large-scale analysis of educator-AI conversations shows that teachers seek actionable guidance and reject outputs that do not align with their instructional context, desired format, or pedagogical purpose (Liu et al., 4 Mar 2025). MathAIde’s design process operationalizes that same principle at the interface level: usefulness is defined not only by technical capability, but by fit with teacher workflow, terminology, and time constraints.

5. Classroom deployment and empirical findings

The classroom case study used the deployed version of MathAIde in real 4th- and 5th-grade classes with 49 students and the same three teachers from the A/B test. Each teacher used the app four times over four days. Every use involved a recommended list of four mathematics exercises, photographing each student’s responses, and correcting MathAIde whenever needed. Across all sessions, 784 student answers were collected. The lists were aligned with Brazilian curricular skills from the BNCC: lists 1 and 2 targeted EF02MA06, and lists 3 and 4 targeted EF03MA07 (Guerino et al., 31 Jul 2025).

The most important quantitative finding is the operational importance of the AuI correction mechanism. Teachers used the correction functionality 139 times out of 784 answers, or 17.6\% of all responses. Of those, 130 answers, or 16.5\%, were corrected by indicating that the student had actually gotten the answer right even though MathAIde had labeled it wrong. Another 9 answers, or 1.1\%, involved changing the error category the AI had assigned. The paper highlights that the need for correction was highly uneven across students; for example, one student, S14, was marked “student got it right” for all 16 captured answers (Guerino et al., 31 Jul 2025).

These results are used by the authors to argue that teacher intervention is not a marginal add-on but a necessary fairness mechanism. Most corrections restored credit to students whose correct work had been misclassified. This is a key empirical rebuttal to the idea that teacher oversight merely fine-tunes edge cases. In this deployment, the human override layer was central to preventing systematic under-crediting of correct handwritten work.

Post-study interviews reinforced that interpretation while surfacing further design tensions. Teachers wanted even more objective and faster correction flows. One teacher argued that the app should focus on helping evaluate whether the student got the answer right or wrong rather than presenting too many detailed error-classification options. Another suggested color coding at the point of image capture, such as green for correct and red or orange for issues, to make review targets easier to detect. The study also found that existing green visual cues could increase trust to the point that teachers skipped detailed checking, indicating that visual design can alter the amount of actual human oversight exercised (Guerino et al., 31 Jul 2025).

This operational human-AI partnership is consistent with evidence from other deployed mathematics-support systems. In VATE, an elementary mathematics system deployed on the Squirrel AI platform, diagnosis from student drafts improved error analysis and downstream learning efficiency, but the design still depended on guided dialogue rather than answer dumping (Xu et al., 2024). In hybrid tutoring settings, human support proved especially valuable for students with lower prior achievement and for contexts where software-only engagement was weak (Thomas et al., 2023). MathAIde’s classroom results extend that broader pattern into handwritten assessment and feedback.

6. Position in the broader math-AI landscape, limitations, and future directions

MathAIde occupies a distinctive position within current mathematical AI systems. It is not a specification-matrix-conditioned exam generator like V-Math, which integrates question generation, solver/explainer functions, and a personalized tutor for a national high-stakes exam (Nguyen et al., 12 Sep 2025). It is not a multi-agent GraphRAG tutoring platform centered on Socratic dialogue and DAG-based course planning (Chudziak et al., 14 Jul 2025). It is also not a handwritten high-stakes grading pipeline that routes uncertain calculus responses through psychometric filters and human review (Kortemeyer et al., 4 Oct 2025). Its specificity lies instead in teacher-mediated capture and correction of handwritten elementary mathematics in low-resource classroom contexts.

Within the broader taxonomy of AI systems for mathematics education, MathAIde combines at least three roles. First, it acts as an information extractor by turning photographed notebook pages into machine-readable judgments. Second, it uses reasoning components to assess correctness and assign provisional error categories. Third, it supports adaptation through feedback, analytics, and recommendations elsewhere in the platform. What it does not attempt to provide in this paper is a deep, formalized explainer layer or a complete student model of the kind sometimes envisioned in broader taxonomies of mathematics education AI (Vaerenbergh et al., 2021).

The paper is also explicit about its limitations. Sample sizes were small, especially in the A/B test and case study. The classroom deployment involved only three teachers and 49 students in localized contexts, so generalizability is limited. The AI remains constrained in scope and accuracy, struggles with the variability of authentic handwritten work, and currently supports only addition, subtraction, and multiplication equations with up to three digits. The deployed correction flow still appears to need refinement in wording, visual signaling, and speed. Long-term adoption and scalability remain open questions (Guerino et al., 31 Jul 2025).

Another important limitation is technical opacity. The paper gives capability statements and performance summaries, but not exact model specifications, OCR architectures, preprocessing pipelines, objective functions, or confidence rules. In that respect, MathAIde differs from tool-augmented reasoning frameworks such as AgentMath, where explicit interaction protocols, reward functions, and training objectives are part of the contribution (Luo et al., 23 Dec 2025). For MathAIde, the primary contribution is design methodology and deployment evidence rather than algorithmic novelty at the model level.

Future work follows directly from these constraints. The authors plan to improve MathAIde’s augmented-intelligence features based on classroom-study insights, especially around faster and more semantically clear review flows. They also intend to conduct larger-scale studies examining effectiveness, efficiency, learning analytics, and learner and teacher experience, and they identify dashboards and teacher training materials as further directions (Guerino et al., 31 Jul 2025).

In synthesis, MathAIde is best understood as a teacher-centered, mobile ITS for handwritten elementary mathematics in which the central innovation is not autonomous grading but a constrained human-AI review mechanism. The evidence suggests that this is not merely a conservative design choice. It is a response to the actual technical and pedagogical conditions of classroom handwriting recognition, low-resource deployment, and fairness-sensitive assessment. In that sense, MathAIde exemplifies a broader lesson emerging across recent mathematics-AI research: in authentic educational settings, the most robust systems are often those that treat AI as a high-leverage component within a carefully designed human-mediated workflow rather than as a complete substitute for educational judgment (Guerino et al., 31 Jul 2025).