A lead on every test
An AI system designed to assess what learners know across several learning domains produced the strongest displayed prediction results across all four multi-domain datasets tested in the preprint. Called LT-MKT, it recorded the highest displayed AUC and ACC and the lowest displayed RMSE in every dataset comparison.
On JuniorH, LT-MKT reported an AUC of 0.9387, an ACC of 0.8425 and an RMSE of 0.1214. On SeniorH, the corresponding figures were 0.9312, 0.8470 and 0.1141. On PTADiscJP, they were 0.7645, 0.8410 and 0.3485; on PTADiscDS, they were 0.6929, 0.8031 and 0.4065.
The benchmarks spanned school subjects and programming. JuniorH and SeniorH included mathematics, physics and English; PTADiscJP covered Java and Python; PTADiscDS covered C programming and Data Structure & Algorithm Analysis. The reported student counts were 1,081, 4,869, 29,430 and 12,271 respectively, alongside 39,230, 133,683, 11,172,165 and 1,788,245 interaction records.
A model built around cross-domain links
The central question was whether making cognitive load and knowledge transfer explicit would improve knowledge-state assessment in multi-domain learning.
LT-MKT uses an LLM to build a Multi-domain Hierarchical Graph from question and concept text. It then models temporal dependencies, relationships across domains and the transfer of knowledge between them.
The study judged next-question prediction with AUC, ACC, or accuracy, and RMSE, an error measure, with the comparison favoring higher AUC and ACC and lower RMSE. It trained the model with binary cross-entropy and split learners at the learner level into training, validation and test sets in an 8:1:1 ratio; validation performance selected the final model.
What the supporting tests found
When parts of the system were removed, the reported scores fell: the full LT-MKT model performed best, while removing cognitive-load modeling, graph-transfer components or state fusion produced lower performance. These were comparisons among model variants, not randomized educational tests, so they do not establish a causal effect in students.
Another test examined the graph used by the model. Expert judgments on a five-point scale gave the GPT-5 pro row an Educational Rationality score of 4.66, Structural Consistency of 4.56 and Fleiss' kappa of 0.84, the highest displayed values in that comparison.
In a representation analysis, the version without cognitive-load modeling was described as highly mixed. LT-MKT showed a clearer, more continuous gradient from low to high cognitive load. The load index, however, was built from model-derived factors rather than direct cognitive measurements.
The preprint also tested a manually constructed cold-start setting in JuniorH, where the test set contained 33 unseen concepts, or 19% of all concepts. Training held 84% of interaction records and testing 16%, and LT-MKT consistently outperformed the other displayed methods on AUC and ACC. The exact plotted AUC and ACC values were not reported in the text excerpt.
Performance was generally best at moderate question-difficulty granularity and around a domain-transition window of 20, according to the sensitivity analysis; that is a preferred range, not a universal optimum.
Where the evidence stops
These results come from an arXiv preprint identified as version 1, dated 25 August 2026. PTADiscJP and PTADiscDS are described as publicly available PTADisc-derived datasets, while JuniorH and SeniorH are described as proprietary iFLYTEK datasets.
The benchmark and ablation results concern model prediction, while the cognitive-load finding concerns a representation produced by the model. The latter was not a direct cognitive measurement, and the ablation comparisons were not randomized causal tests.
The paper states that implementation code is available in the LT-MKT GitHub repository.
Paper data and sources
Original title: Incorporating Cognitive Load and Knowledge Transfer for Multi-Domain Knowledge Tracing
Authors: Haotian Zhang, Shucun Wang, Jinze Wu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text