Personalisation papers lack mechanics

The missing piece in many artificial intelligence learning papers is not the promise of personalization but the rule that decides what a student sees next. (ijctjournal.org) A personalized system usually does two jobs. It first estimates a learner’s current mastery from past answers, a process researchers call knowledge tracing, and then chooses the next lesson, hint, or question from that estimate. (educationaldatamining.org) The paper hosted by the International Journal of Creative Research Thoughts describes a platform called PERSONA and says it adapts content in real time, profiles users, recommends material, and gives immediate feedback. The abstract does not spell out the state model, training data, or decision policy used to make those adaptations. (ijctjournal.org) That omission stands out because the mechanics are the system. In adaptive learning, the estimate of what a student knows can come from simple probability tables such as Bayesian Knowledge Tracing or from neural models such as Deep Knowledge Tracing, and each choice changes how the platform behaves. (nature.com) (educationaldatamining.org) The second job is the recommendation policy, which is the rule for picking the next task once the system has a guess about mastery. Researchers often frame that step as reinforcement learning, where software tries actions, measures rewards such as accuracy or completion time, and updates the policy over many interactions. (educationaldatamining.org) (nature.com) A November 17, 2025, Scientific Reports paper shows what that level of detail looks like. It says its RL-DKT system combines Dynamic Knowledge Tracing with Reinforcement Learning, tests on ASSISTments, KDD Cup 2010, and Cognitive Tutor, and reports a 12.5 percent improvement in task completion time, a 50 percent drop in dropout, and a 7.6 percent gain in prediction accuracy over baseline models. (nature.com) Even then, strong benchmark scores do not settle deployment questions. A 2025 Khan Academy industry paper using data from more than 500,000 students and 100 million interactions found practical weaknesses in deep knowledge tracing models, including cold-start problems for new students, weaker detection of incorrect responses, and sensitivity to question order. (educationaldatamining.org) That is why summaries that say a system finds knowledge gaps and adjusts difficulty leave out the part teachers and buyers need most. Without a clear account of how mastery is inferred, what objective is optimized, and how recommendations are constrained, “personalized” can describe anything from a fixed rules engine to a fully learned policy. (ijctjournal.org) (educationaldatamining.org) The parallel MarketingProfs roundup published April 3, 2026, follows the same pattern at a higher altitude. It lists broad artificial intelligence product moves by OpenAI, Microsoft, Salesforce, and Anthropic, but it does not add operational detail about how educational adaptation systems estimate learner state or choose interventions. (marketingprofs.com) For readers trying to assess personalized learning claims, the basic questions are plain ones: what does the model think the student knows, what evidence updates that belief, and what rule picks the next step. If a paper does not answer those three questions, it is describing the goal more than the machine. (educationaldatamining.org 1) (educationaldatamining.org 2)