PersonaPath: A Knowledge-Centric Benchmark for Personalized Learning Path Planning
Synopsis
This work introduces Knowledge-Centric (KC) personalized learning path planning and builds PersonaPath, a benchmark pairing 2,000 fine-grained learner personas with a hierarchical knowledge graph of 347 textbooks, 1,751 units, and 4,092 concepts across 77 subjects, to evaluate whether large language models can decide which textbook, unit, and concept a learner should study next given learner profiles, mastery states, and prerequisite knowledge structures; evaluation of representative LLMs shows the strongest model reaches only a 29.5% final pass rate in Basic Education, with adaptivity as the main bottleneck, where no model exceeds 44.7% in tailoring paths to individual learners.
Figure 1: Illustration of Exercise-Centric (EC) recommendation and Knowledge-Centric (KC) planning. Panel A illustrates recommendations conditioned on correctness records alone; Panel B additionally uses explicit learner goals and mastery states. Each node (E1, E2, …) represents an exercise item, and edges denote prerequisite relations among the knowledge points. Abbreviations: VC = Vector Coordinates, VA = Vector Addition, DP = Dot Product, VD = Vector Decomposition, CA = Component-wise Addition, and CDP = Component-wise Dot Product.
arXivInterpretation
It formulates Knowledge-Centric (KC) personalized learning path planning, distinct from exercise-centric recommendation. Adaptive learning systems commonly frame path planning as Exercise-Centric (EC) recommendation, inferring the next step from item-level interaction logs; this work argues that goal-oriented guidance additionally requires explicit learner goals and curriculum-scale prerequisites, since learners with similar exercise records may need different paths toward their targets. A task-definition contribution; the abstract explicitly contrasts the EC and KC settings, so this is a conceptual and problem-setting contribution rather than an experimental effect size.
It constructs the PersonaPath benchmark, pairing fine-grained learner personas with a hierarchical knowledge graph. The benchmark contains 2,000 fine-grained learner personas and a hierarchical knowledge graph spanning 347 textbooks, 1,751 units, and 4,092 concepts across 77 subjects, providing a resource for curriculum-scale, goal-oriented path planning evaluation. The resource scale is stated explicitly in the abstract (2,000 personas; 347 textbooks, 1,751 units, 4,092 concepts, 77 subjects), making it a checkable description of the benchmark's composition.
Evaluation of representative LLMs shows limited current performance on this task, with the bottleneck concentrated in adaptivity. Results show that even the strongest LLM reaches only a 29.5% final pass rate in Basic Education, and that no model exceeds 44.7% in tailoring paths to individual learners, pointing improvement efforts toward individualization rather than knowledge coverage alone. These are evaluation numbers reported in the paper (29.5% final pass rate, 44.7% adaptivity ceiling), representing model comparisons on the benchmark; the specific model list, evaluation protocol, and statistical details require consulting the original text.
Perspective
This work targets the knowledge-centric learning path planning setting, applicable where explicit learning goals, learner profiles, and curriculum-scale prerequisite knowledge structures exist, such as subject instruction organized by textbook, unit, and concept hierarchies; its benchmark covers 77 subjects and includes evaluation in the Basic Education setting, so the conclusions mainly apply to such curriculum-structured, goal-oriented guidance tasks rather than open-ended exploratory learning without explicit goals or pure exercise recommendation.
The currently available text is the abstract and metadata, lacking the model list, prompt design, evaluation protocol, scoring rubric, and ablation analyses from the body, so it is not possible to judge how the 29.5% and 44.7% figures distribute across models, subjects, and difficulty levels; additionally, the provenance of the personas and knowledge graph, annotation consistency, and the benchmark's external validity in real instruction still require confirmation from the original text.