Research at Aristotle
Building AI that helps people think for themselves
A tutor’s success is what the student learns. Our research starts there—and asks how to make every conversation serve that goal.
Why this is worth researching
Building a good AI tutor is hard.
An AI tutor can answer questions without the student learning a single thing. This is a byproduct of the way AI systems are built. Big labs spend millions on training to make their AI systems helpful assistants that maximize productivity. And while these systems are climbing benchmarks to be world-class mathematicians, they often fail to be effective at teaching fractions. Why is that?
Four challenges in one
How our research fits together
Define what great tutoring looks like
What should a tutor do next?
The same hint can help one learner and do too much of the work for another. A good tutor has to notice what the student understands, decide when to explain, and know when to leave room for them to think.
- Session leadership
What is worth working on next, and when should we move on?
- Cognitive load
Is the student doing the thinking, with enough support to keep going?
- Communication
Does this explanation make sense to this particular learner?
- Accuracy
Are the ideas, calculations, and feedback correct?
Measure tutoring with human judgment
Would an experienced tutor agree?
A correct answer is only one part of a good tutoring exchange. We use examples reviewed by human experts to develop shared criteria, then align automated evaluators with those judgments.
Compare expert judgments
Review the same exchanges and work through where human reviewers disagree.
Align automated evaluators
Use those expert-reviewed examples to evaluate teaching across more sessions.
Examine the reasons
Check whether the evaluator recognized the teaching decision behind its score.
Simulate students to study whole sessions
Can a simulated student respond like a real learner?
We need to develop simulated students to test how a tutor’s decisions play out over a conversation. Rather than grading a single reply against a fixed transcript, we repeatedly run new versions of our tutor against these simulations. That lets us study patterns of teaching that only emerge across a session.
Sounding like a student is only the start
A model playing a student may solve a problem it was supposed to misunderstand, or follow an explanation that would leave a real learner confused. If the simulated student learns too easily, a weak tutor can look effective.
Start with a learning situation
Ground the scenario in what a student knows, where they’re stuck, and examples from real tutoring.
Let the exchange unfold
Have the tutor and simulated student respond to each other across multiple turns, so the tutor encounters the consequences of its own decisions.
Evaluate the teaching—and the simulation
Apply our teaching criteria to the exchange while examining whether the student’s responses and changes in understanding are believable.
Measure what students actually learn
Can the student do more on their own?
Better teaching scores matter when they lead to better learning. We’re developing ways to connect what happens during tutoring to what a student can understand and solve independently, within a session and across the sessions that follow.
From supported practice to independent understanding
Find the starting point
Notice what the student can do before receiving help.
Track the support
Track the support they need while working through a problem.
Check independent understanding
Return to the skill with a different problem and see what they can do on their own.
With support
Work through 3x + 7 = 19 together.
On their own
Try 4x + 5 = 21. Explain the first step and why it keeps the equation balanced.
responsible research
Our students come first
Our researchers and product team work together: questions from tutoring inform our research, and what we learn helps us improve the tutor. We use internal evaluation and student simulation to investigate improvements to our tutor, and we only release them to real students if we think they'll help.
That care extends beyond teaching quality. Our safety approach brings together checks during tutoring, session review, and human judgment. Learn how we respond to concerns and handle your family’s information.
Our approach to safety
if you didn't need a TL;DR...
Join us to build a tutor worth learning from
How do you know when a hint helps? What shows that a student understands? Join us to turn questions like these into research, experiments, and better teaching.
From the notebook
A closer look at our teaching decisions and the methods we use to study them.
