Backed by decades of learning science

Students don't just learn from AI.
They learn by teaching it.

MELT-Math is a teachable AI agent for mathematics. Instead of handing out answers, it becomes the learner — asking questions, making mistakes, and prompting students to reason, reflect, and truly understand.

150,000+
messages analyzed
60,000+
students reached
1.5
effect size across classroom studies
The Problem

AI that just gives answers makes students passive.

Today's AI tutors optimize for the fastest correct answer. But when the machine does the thinking, the student stops. Understanding stays shallow and fades fast.

Traditional AI tutor

How do I solve 2x + 6 = 14?

Sure! Subtract 6 from both sides → 2x = 8. Divide by 2 → x = 4. Done. ✅

Student copies the answer and moves on. Nothing sticks.

Passive learning

Students receive answers instead of constructing their own reasoning.

Reduced agency

Students rely on AI rather than making decisions and explaining their thinking.

Weak ownership

Learning feels like following instructions instead of building understanding.

Our Solution

Meet MELT-Math — the AI that learns from your students.

We flip the relationship. The student becomes the teacher, and the AI becomes a curious, fallible learner. Explaining your reasoning is where deep understanding is built.

  • The student leads — they explain, the AI listens
  • The AI asks for clarification instead of giving answers
  • Hidden misconceptions surface as the student teaches

A negative times a negative gives a negative, so (-3)(-4) = -12.

Got it — so two negatives stay negative? Let me try (-5)(-1) = -1 then. 🤔

AI makes a deliberate error

Wait, no — I was wrong! Two negatives cancel, so (-3)(-4) = +12.

Ahh, so the signs flip to positive. Can you show me more examples?

Student catches it & re-explains
How learning-by-teaching works

A simple loop that builds deep understanding.

Three steps, repeated. Each cycle pushes students from recall toward real reasoning.

01

Learn

Students explore a concept with curated, multimodal material — notation, diagrams, and visual reasoning.

02

Teach

They explain it back to MELT-Math, which plays the curious learner — asking questions and probing the 'why'.

03

Reflect

Misconceptions surface, the student rethinks and refines, and understanding gets reinforced through reflection.

Why we're different

Not another answer engine.

Typical AI tutors

Optimized to answer

  • Delivers the solution on demand
  • Student stays passive
  • Rewards copying over thinking
  • Understanding fades quickly
  • One-size-fits-all explanations

MELT-Math

Optimized to make you think

  • Asks questions instead of answering
  • Student actively teaches & reasons
  • Rewards explanation and reflection
  • Builds durable, transferable understanding
  • Adapts to each student's misconceptions
Core Features

Everything a teachable agent needs.

A complete platform engineered around the science of how students actually learn mathematics.

Teachable AI Agent

An AI that plays the learner — curious, fallible, and always nudging students to explain their reasoning.

Multimodal Mathematics

Work in natural language, formal notation, diagrams, and visual reasoning — all in one fluid session.

Mentor Agent

A supportive AI mentor that provides timely hints, asks guiding questions, and helps students move forward without taking over the thinking.

Misconception Detection

The agent surfaces and gently challenges hidden errors in a student's mental model.

Adaptive Feedback

Responses adjust in real time to each student's level, pace, and reasoning style.

Research-Based Design

Every interaction is grounded in decades of learning-sciences and learning-by-teaching research.

Research Foundation

Built on the science of learning.

MELT-Math isn't a hunch — it's the product of decades of peer-reviewed research into how people learn, reason, and retain.

  • Learning by Teaching

    The protégé effect: students who teach a peer — or an agent — learn more deeply than those who study to perform.

  • Metacognition

    Reflecting on one's own thinking is one of the strongest known drivers of durable learning.

  • Constructivist Learning

    Knowledge is actively constructed by the learner, not passively transmitted from a source.

  • Learning Sciences

    Modern findings on feedback, spacing, and retrieval shape every interaction in the product.

Public Good

Research translated into public good.

MELT-Math is not only published and peer-reviewed. It is built as shared educational infrastructure: open models, reusable datasets, responsible data-sharing agreements, and classroom partnerships reaching 60,000+ students across five school districts.

60,000+

students reached

Across five partner school districts.

5

school districts

Classroom partnerships supporting real-world implementation.

Open

models and code

Reusable AI models and codebases shared with the field.

Shared

datasets and agreements

Research datasets supported by responsible data-sharing pathways.

Evidence Base

Peer-reviewed research remains part of the public-good story.

Publications document the learning science behind MELT-Math, while open models, datasets, and district partnerships make that work reusable beyond a single study.

Who it's for

Made for everyone who shapes learning.

Students

Build genuine understanding and confidence by teaching what you're learning.

Deeper mastery, less cramming

Teachers

Get visibility into how students reason — and where misconceptions hide.

Insight into real understanding

Schools

Deploy an evidence-based tool that lifts outcomes across classrooms at scale.

Measurable, district-wide impact

Researchers

A rich platform to study metacognition and learning-by-teaching in the wild.

New data on how students learn
From the classroom

Teachers are watching students become the teacher.

Real reactions from educators in MELT-Math classroom pilot studies.

95% of learning is done through teaching. I love how the kids are actually able to type in and explain — it takes the teacher away from being the standard deliverer and makes all the kids little teachers themselves.

6th Grade Math Teacher

MELT-Math Pilot Study

It's really good for kids to practice explaining their answers — that's something we don't always get to hear from every kid. It would be nice for all of them to have that opportunity.

7th Grade Math Teacher

MELT-Math Pilot Study

I've never had a program that encourages students to really explain their answers, and that's important.

7th Grade Math Teacher

MELT-Math Pilot Study

The students feel very excited when I tell them we're going to use this math tool — and they love that they're the ones being tutored instead of the reverse.

8th Grade Math Teacher

MELT-Math Pilot Study

On using MELT-Math again
I'd love to try it again, especially with my lower-level students. It could really challenge them to understand the why behind math.

8th Grade Math Teacher

MELT-Math Pilot Study

Future Vision

A new default for how students learn.

We're starting with mathematics — but learning by teaching is universal. Here's where MELT-Math is going.

Beyond mathematics

Extend the teachable-agent model to science, language, and reasoning across the curriculum.

Every classroom

Multilingual, accessible, and affordable — so any student anywhere can become the teacher.

Teacher superpowers

Live dashboards that reveal class-wide understanding and misconceptions as they form.

Now partnering with districts & schools

Join us in transforming mathematics learning.

See MELT-Math in action and discover what happens when students teach the AI.