Tutoring apps that adapt to each learner run into a catch: you rarely have much data on any single person.

This explores why having only a little data about each individual learner makes it hard to build tutoring systems that adapt to them. The corpus has no notes on tutoring or human education, so what follows is a lateral read from neighboring work on personalization and teacher-student mismatch. It is not a direct answer.

The first neighbor is the closest thing to a fix. One line of work suggests the bottleneck is less 'too little data per person' than 'asking the wrong questions.' Can user preferences be learned from just ten questions? learns a shared set of base preferences from many people, then pins down one person's position among them using about ten adaptively chosen questions, with no retraining. It tackles preferences over AI responses, not what a student knows, so the transfer to tutoring is untested. But it shows the shape of the answer: pool the population to learn the structure, then spend a small per-person budget on the questions that reduce uncertainty most.

The second neighbor explains why per-student data matters so much. Does teacher-refined data always improve student model performance? finds that a better explanation is not always a better lesson. Teacher-refined data hurt when it went beyond the student model's learning frontier, and students did better filtering refinements against their own profile. That is a tutoring problem in miniature. You can't pick the right level of explanation without knowing where the learner is, and that is exactly what sparse data fails to tell you. Does richer teacher context hurt student generalization? adds a second risk. Teachers who see the answer produce confident, tidy traces, and students inherit that confidence and generalize worse to unfamiliar problems. A tutor that guesses wrong about a learner may not just be unhelpful. It may teach false certainty.

A third cluster shows what happens when the signal is weak, and how to pick better signal. Why do language models collapse into generic templates? shows that when feedback barely differs across attempts, the model drifts toward generic, input-ignoring answers. Filtering for high-variance, informative prompts fixes it. A tutor with a thin history on a student is in the same position, since nothing distinguishes this learner from the average and the easy outcome is a one-size-fits-all lesson. Yet Can a single training example unlock mathematical reasoning? shows that a single well-chosen example can lift math accuracy from 36% to 73.6%. Thin data is not always fatal if it activates something already latent. Together, these two suggest the quality and contrast of each data point matter more than the count.

A looser idea, and my extrapolation rather than something the notes test on students: some systems can read difficulty from their own internals instead of from labeled histories. Can representation sparsity order few-shot demonstrations effectively? orders examples from hard to easy using activation sparsity, with no external difficulty labels. Can simple uncertainty estimates beat complex adaptive retrieval? finds a model's own uncertainty beats elaborate external heuristics for deciding when to fetch help. Those signals describe the model, not the learner. But they hint that an adaptive tutor might lean on cheap self-generated signals to compensate for the missing per-student record. The corpus here is silent on whether that works for real human learners.