MathGym for Educators

A student may have mentioned MathGym, a colleague may have flagged it, or you may be looking for a study tool you could actually endorse. Either way, this page tells you who built it, how it works, exactly where the AI is and isn't, what it costs, and what it will never do.

What MathGym is, in one minute

It has never been easier for a stuck student to get a solution. ChatGPT, Gemini, and the rest will produce a detailed step-by-step answer to almost any homework, review sheet, or practice exam in seconds. You've seen where that leads, and we have, too: many students did super well on the homework sets, yet they struggle on the exam.

The catch is old news to the math educators: students do not learn math by reading solutions. In fact, long before AI, students followed examples, copied patterns, and leaned on outside help for homework without ever building the underlying skills. AI just made it easier for students to fall into this trap.

MathGym is our answer to the problem. Our assumption is that when a student can't solve a problem, they're missing a specific skill, and often a prerequisite one. Here is a quick summary of how MathGym works.

• Upload: The student uploads a math problem (photo, screenshot, or typed).
• Skill Identification: MathGym identifies the specific underlying skills the problem requires, including the prerequisite skills (eg., trigonometry for a calculus problem). Each skill has its own set of practice problems, which are randomized variants, each with fully worked solutions. They are verified by math educators, and every answer is deterministically graded by a grading code, not by AI.
• Targeted practice. Based on the identified skills, MathGym generates a practice session.
• Coaching, not answers. A built-in AI coach offers hints one at a time and is designed never to give away the solution.
• Test Prepared: Through deliberate practice, your student builds the specific skills needed to solve the original problem and walks into the quiz or exam more prepared and confident.

What's human-verified, and where the AI is

Human-verified (the mathematical content):

• The problem bank. Every practice problem comes from a bank of thousands of problem generators, each producing a number of randomized variants of a problem type with complete solutions. Two working mathematicians review and verify these generators before they reach a student. And the problem a student practices on is never invented on the spot: it comes from a fixed, pre-verified generator, not from a model improvising a question and answer in real time. We are human beings and can make mistakes. However, if a problem is ever wrong, it's a bug in a specific generator. We fix that generator once, and it's permanently corrected for every student after.
• The grading. Every answer is graded by deterministic grading code. The same input always gets the same verdict. No language model ever judges whether a student's mathematics is correct.
• The pedagogy. The skill taxonomy and the practice structure stem from our teaching experience.

AI, in the live session (the plumbing and the coach):

• Reading the upload. A model reads the student's uploaded problem and maps it onto our skills taxonomy.
• The hint coach. A guardrailed chat coach that offers hints, one per turn, and is built to never give the solution.

No Learning Shortcuts

We know exactly what a stuck, tired student does at 1am, and we know what it costs them in Exams. MathGym's guardrails exist because AI answer-machines hurt learning. This is why this product exists.

What that means concretely:

• There is no answer to copy. The coach gives hints, one per turn, and is designed never to state the final answer or a transcribable solution. A student cannot paste your assignment into MathGym and walk away with something to submit.
• Practice problems are not your assignment. Generated problems are randomized variants that train the underlying skills. Full solutions are also hidden from students for these problems. They can ask AI tutor for hints, but the tutor will NOT just give the students answers or solutions.
• The incentive structure points to exams. Everything in the session is designed to help the student perform independently on quizzes or exams.

If you have an integrity concern we haven't addressed here, please email us. Your opinion and suggestions are crucial to us as we improve our product to promote learning.

Why not just let AI generate the practice problems?

It's a fair question. If a chatbot can write out full solutions, why not let it write the practice problems too? We tried it. Two problems make it a bad trade.

1. The problem with accuracy

When a standard AI generates a math problem on the go, it invents the question and the answer simultaneously. Having spent a decade writing homework and exams, we know that one of the trickiest parts of designing a good problem is the numbers.

Good problems use carefully chosen figures so students can focus on the underlying concept rather than getting bogged down in messy, distracting arithmetic. This is particularly difficult for large language models.

Graph-based problems are worse: reading or building a question from a visual graph is where today's models stumble most, burning a large amount of tokens for a result you still can't trust. What's worse is that an incorrect AI problem vanishes the moment a student moves on, so any hidden structural flaw just rolls into the next session.

2. The problem with control

AI grading is not a definitive check. Standard AI can mark a perfectly correct answer wrong simply because it was written in a different format. Furthermore, without strict boundaries, randomly generated AI problems can drift away from the specific target skill and difficulty level the student actually needs.

The MathGym difference: engineered precision + responsible use of AI

MathGym's problems come from roughly thousands of generators built by mathematicians, each focused on a specific micro-skill.

• Deterministic code grading. Every answer is checked by precise underlying code that tests for true mathematical equivalence.
• Permanent fixes. We're human, and we make mistakes. But unlike a fleeting AI hallucination, if a bug is found in our problem bank, we fix the generator once, and it is permanently corrected for every student going forward.
• Responsible use of AI. We keep AI completely out of writing problems and grading math. We use it exactly where it's strong: reading the student's uploaded problem, mapping the underlying skills, and acting as a supportive coach that delivers hints one step at a time.

Quick version:

Who we are

MathGym is built by its two founders, both former research mathematicians. There is no growth team, no sales team, and no outsourced content shop. When you email MathGym, one of us answers.

• Samuel Lin, Ph.D. — Ph.D. in mathematics, Michigan State University. Taught at Dartmouth College, the University of Oklahoma, and Colby College. Published research in Geometric and Functional Analysis (GAFA), Transactions of the AMS, and Geometry & Topology.
• Wenchuan Tian, Ph.D. — Ph.D. in mathematics, Michigan State University. Former mathematics instructor at UC Santa Barbara. Published research in Mathematische Annalen and Calculus of Variations and PDE.

What MathGym costs

MathGym is free during beta — full access, invite-based. Beta access currently means 40 problem uploads a month (each generating a complete targeted practice session) and generous daily coaching limits. There is no way to accidentally spend money on MathGym today: we don't collect any payment information.

What we never do

• We never sell personal data.
• We never do a student's homework. The coach doesn't give answers; the practice problems are randomized variants (not a lookup table for any course's assignments), and the session ends with the student solving their own problem themselves.
• We never bill silently.
• We never claim to replace teachers or tutors. MathGym is a study tool that provides practice and coaching, but it is not a replacement for reading the textbook, attending lectures, or attending office hours.

How your students get started

MathGym is in an invite-based beta. Students can join the waitlist at mathgym.ai/signup and will receive a response within 2 days. The account is free, and no payment information is required at any point in the process.

Educator Account

If you are interested in trying MathGym, please email us to sign up for a free educator account. If you would like to explore whether MathGym is suitable for your institution, please email us, and we can schedule a 20-minute personal walkthrough with a founder.