A coaching institute in Kota runs a batch of 90 students through the same JEE problem set every evening. Roughly a dozen of them finished the set in twenty minutes and are bored. Another dozen never got past question four and have quietly stopped trying. The remaining sixty are somewhere in between. One teacher, one pace, one outcome distribution that has been stable for a decade.
That is the problem adaptive tutoring is aimed at, and after several years of overpromising, the technology has become good enough that Indian institutes are running it in production rather than in pilots.
What a Learning Path Really Is
A personalised learning path is not a playlist. It is a model of what a student knows, updated after every interaction, plus a policy for choosing what comes next.
The knowledge model is usually some variant of knowledge tracing. Classical Bayesian Knowledge Tracing represents each skill with four parameters: prior knowledge, learning rate, slip probability, and guess probability. Deep knowledge tracing replaces that with a sequence model over the student’s full attempt history. Both are trying to answer the same question, which is the probability this student gets the next item on this skill correct.
The selection policy then targets a difficulty band. The research consensus sits around a 70 to 85 percent success rate as the productive zone. Below that students disengage, above it they coast. If your system is serving items a student answers correctly 95 percent of the time, it is not tutoring, it is padding a completion bar.
Where LLMs Add Value and Where They Do Not
Large language models have changed the explanation layer far more than the sequencing layer.
- Worked solutions on demand are genuinely better. A model can rewrite the same integration-by-parts explanation four ways until one lands, which no printed key can do.
- Socratic hinting works when constrained. Systems that are instructed to ask a guiding question rather than reveal the answer produce better retention than systems that just show the solution.
- Misconception diagnosis from free-text answers is now practical. A student who writes that dividing by a fraction makes a number smaller has a nameable, addressable error, and a model can catch that from prose where a multiple choice item cannot.
- Item sequencing is not an LLM strength. Ordering content by mastery probability is a statistics problem, and a tuned knowledge tracing model beats a prompt every time.
- Grading high-stakes work should stay out of the model’s hands without human review. Agreement rates with expert graders on open-ended science answers still run well short of what an exam result requires.
Making It Survive Contact With a Real Classroom
Adaptive systems fail in Indian institutes for reasons that have nothing to do with model quality. The most common one is that the syllabus is fixed by a board or an entrance exam and the teacher cannot let the system wander. Any path engine that cannot be pinned to a chapter sequence will be switched off by February.
The second is teacher visibility. If a faculty member cannot open one screen and see which five students in a batch of 90 are behind on which specific topic, the tool is invisible to the person who decides whether it gets used. The third is cold start. A new student in week one has no history, and serving them random difficulty for three sessions is enough to lose them. A short diagnostic, ten to fifteen calibrated items, buys a usable prior immediately.
Finally, treat mastery decay as real. A topic mastered in June is not mastered in November. Spaced review scheduled against a forgetting curve is often the single highest return feature in the whole system, and it is far cheaper to build than the tutoring layer.
At Invexa, we build learning systems where the knowledge model, the teacher dashboard, and the syllabus structure are designed together, because an adaptive engine that a faculty member cannot read or override does not survive its first term.