TLDR: V-MATH is an autonomous AI framework designed to assist Vietnamese high school students and teachers with the National High School Graduation Mathematics Exams. It features three specialized agents: a question generator, a high-accuracy solver/explainer, and a personalized tutor. The system leverages a multi-tier memory architecture, including a Memento-style case memory for continuous learning, and has demonstrated superior performance in generating compliant, novel exam questions, providing accurate step-by-step solutions, and customizing learning paths, significantly improving student outcomes and reducing teacher workload.
Preparing for high-stakes examinations can be a daunting task for students and a demanding one for educators. In Vietnam, the National High School Graduation Mathematics Exams (NHS-GMEs) present significant challenges, requiring students to master diverse topics across various cognitive levels. Traditional methods often fall short in providing personalized, adaptive, and creative learning experiences. This is where V-MATH, an innovative autonomous agentic framework, steps in to transform mathematics exam preparation.
Introducing V-MATH: An AI-Powered Learning Companion
Developed by researchers from HUTECH University and other institutions, V-MATH is designed to assist Vietnamese high school students in their journey towards excelling in the NHS-GMEs. More than just a study tool, it’s a comprehensive system that integrates specialized artificial intelligence agents to offer a dynamic and personalized learning environment. The framework also extends its support to teachers, helping them generate high-quality, compliant exam questions and build diverse question banks, thereby reducing their manual workload and enriching instructional resources.
The Brains Behind V-MATH: A Multi-Agent Architecture
At its core, V-MATH operates on a sophisticated multi-agent architecture, powered by the Gemini 2.5 Pro large language model. This system is structured as a planner-executor framework, ensuring iterative task completion and continuous improvement. It comprises a central Planner Agent and three specialized Executor Agents:
The Planner Agent: This high-level AI acts as the strategist, receiving requests from students or teachers and breaking them down into manageable sub-tasks. It makes crucial decisions, such as identifying a student’s knowledge gaps or formulating personalized learning plans.
Agent 1: Creative Exam Generation: This agent is responsible for creating novel exam questions that strictly adhere to the Vietnamese Ministry of Education’s specification matrix. It ensures that questions cover the right topics, sections, and difficulty levels, preventing repetition and fostering originality. It generates full exam sets, complete with multiple-choice options, detailed step-by-step solutions, and clear explanations.
Agent 2: High-Accuracy Exam Solver and Guidance: When presented with an exam paper, whether in PDF format or as image-based inputs, this agent springs into action. It extracts content, converts it into a structured format, and uses advanced computer vision, symbolic reasoning, and natural language processing to solve problems. Crucially, it provides not just the correct answers but also comprehensive, step-by-step explanations, often including alternative methods and highlighting common pitfalls.
Agent 3: Personalized Learning Path Customization: This agent is the student’s personal tutor. It analyzes a student’s performance data, identifies specific skill gaps, and then recommends adaptive study plans. It can generate targeted exercises, customized practice scenarios (like timed drills), and even simulate full-length exams, ensuring that learning is always relevant and tailored to individual needs.
Learning from Experience: The Memento Integration
A key feature enabling V-MATH’s continuous adaptation and improvement is its integration of the Memento framework. This allows the system to learn from diverse student interactions and evolving curriculum demands without the need for constant, costly fine-tuning of the underlying AI models. By storing past successful and failed experiences in a ‘Case Bank,’ V-MATH can recall and adapt its strategies, making its personalized recommendations and question generation even more effective over time.
Rigorous Evaluation and Impressive Results
The effectiveness of V-MATH was rigorously tested using two datasets: the newly curated NHSGMEs dataset (500 complete exam sets aligned with the latest 2025 MOET format) and the public VNHSGE dataset (250 questions from 2019-2023). The results were remarkable:
V-MATH achieved a perfect 100% exact-match accuracy on the VNHSGE dataset across all years (2019-2023), significantly outperforming other powerful models like o1-preview (87.6%) and GPT-4 Omni (86.4%).
On the more comprehensive NHSGME exams, V-MATH demonstrated high section-wise accuracies: 98.1% for Recognition, 93.8% for Comprehension, and 88.4% for Application questions.
The system’s ability to generate creative and compliant questions was validated with a 96.7% matrix compliance rate and a high teacher rating of 4.6 out of 5.
For personalized learning, students following V-MATH’s recommendations showed an average improvement of 11.8 points in their scores, with a 43% reduction in repeated skill errors.
These findings underscore V-MATH’s potential to provide scalable, equitable mathematics preparation aligned with national standards, while also empowering teachers through AI-assisted exam creation. For more technical details, you can refer to the original research paper: V-MATH: An Agentic Approach to the Vietnamese National High School Graduation Mathematics Exams.
Also Read:
- Empowering Students with a Local GPU-Accelerated AI Assistant
- VideoAgent: Crafting Personalized Scientific Videos from Research Papers
Looking Ahead
While V-MATH has already shown significant promise, future work aims to further enhance its capabilities by reducing latency, improving robustness to noisy inputs, expanding teacher controls, and extending its application to other subjects and educational levels. The goal is to create a trustworthy and impactful AI ecosystem for education.
