Ekdant

The best AI tutoring systems in the world are trained by rewriting the entire model. Take a capable language model — one that can reason through math, explain concepts, hold a conversation — and fine-tune all of it with reinforcement learning to make it a better teacher. This is how Google's LearnLM works. This is how the leading academic approach out of ETH Zurich operates.

There's a catch. When you retrain an entire model to improve its teaching, you risk breaking the very capabilities that made it useful in the first place. The model gets better at choosing when to give a hint versus a worked example, but it might get worse at the math it's supposed to be teaching. The technical term is catastrophic forgetting. The plain-language version: you're improving the pedagogy while degrading the intelligence underneath it.

And it's expensive. Reinforcement learning over billions of parameters requires serious compute. That means only well-funded labs and large tech companies can build AI tutors at the current quality bar. The students who need AI tutoring most — in under-resourced schools, on limited hardware, in regions without access to cloud infrastructure — are precisely the ones these approaches can't reach.

We think understanding what a student said and deciding what to do next as a teacher are fundamentally different skills. A tutor needs to know math. But the decision of whether to give a hint, ask a probing question, or walk through an example is a teaching skill, not a math skill. Current systems tangle these together. We're exploring architectures that keep them apart.

The practical consequence we're after: a teaching strategy that's cheap to train, doesn't degrade the underlying model's reasoning, and is portable across models of different sizes. Train the teaching once. Deploy it on whatever hardware a school actually has.

This opens questions we find genuinely interesting:

We're testing in math tutoring first, with a direct comparison against the current state-of-the-art that retrains the entire model. Same problems, same simulated students, same evaluation — the only variable is where the teaching strategy lives.