Incorporating Cognitive Load and Knowledge Transfer for Multi-Domain Knowledge Tracing

Synthesis: Knowledge tracing aims to assess students' dynamic knowledge states from their learning histories, but most existing methods focus on single-domain learning despite real-world scenarios involving multiple domains simultaneously. Zhang and colleagues introduce two critical factors for multi-domain settings: cognitive load arising from managing learning across domains in temporal and knowledge dimensions, and knowledge transfer where knowledge states in one domain influence related states both within and across domains. Their proposed LT-MKT method integrates textual information from questions and their associations to bridge isolated domains, improving knowledge state assessment in multi-domain learning scenarios.

Key Findings

1. Real-world learning often involves multiple domains simultaneously, introducing cognitive load and knowledge transfer as critical factors for Knowledge Tracing.
2. Cognitive load arises from managing learning across domains in both temporal and knowledge dimensions.
3. Knowledge transfer captures how knowledge states in one domain influence related states both within and across domains.
4. LT-MKT integrates textual information from questions and their associations to bridge isolated domains.
5. The method improves students' dynamic knowledge state assessment in multi-domain learning scenarios.

What this means for practice

• Designers. Model cognitive load explicitly in multi-domain knowledge tracing — question difficulty, domain transition, and domain coverage — instead of optimizing only for single-domain sequences.
• Designers. Use Large Language Models (LLMs)-constructed multi-domain hierarchical graphs to link otherwise isolated domains, so that cross-domain knowledge transfer informs each student's estimated knowledge state.
• Designers. Expect returns from stronger LLM backbones: graph-construction quality, and with it tracing performance, improved as the backbone model improved.
• Researchers. Benchmark multi-domain models against cross-domain baselines — students switched domains frequently (average cross-domain interactions of 8.72–14.26 per learner across datasets), so single-domain KT benchmarks understate the problem.
• Researchers. Treat the three load proxies (difficulty, transition, coverage) as an initial set, since the authors explicitly leave other contributors to cognitive load for future work.

Limitations

• Evaluation is offline response prediction on four datasets (JuniorH with 1,081 students, SeniorH with 4,869, and the PTADiscJP and PTADiscDS splits), not a classroom intervention; no teaching or learning outcome is measured.
• Performance is judged only by prediction accuracy against 11 baselines on an 8:1:1 learner split, so better prediction is not itself evidence that modeling cognitive load improves instruction.
• "Cognitive load" is operationalized through three proxy features, and the authors state that other influencing factors are left to future work — no subjective, behavioral, or physiological load measure is used to validate the construct.
• Graph construction depends on an external LLM backbone (Qwen-plus) and its output; performance therefore varies with that model, and the effect of its construction errors on tracing accuracy is not analyzed.

Citation

Zhang, Wang, Wu, Ding, Liu, Huang, Sha, Wang, & Liu (2026). Incorporating Cognitive Load and Knowledge Transfer for Multi-Domain Knowledge Tracing.