📄 CAKT-Agent：基于MoE的情境自适应知识追踪智能体

Summary

Knowledge tracing (KT) estimates whether a learner will answer the next exercise correctly, based on their earlier interactions. The authors argue that current deep KT models have three weaknesses. They reduce learning to one feature, they handle varied learning situations poorly, and they output bare probabilities that teachers cannot act on.

CAKT-Agent addresses this with three cooperating parts:

• Unified memory module. It stores interaction logs, educational rules and cached model states, so the model does not have to recompute everything for each request.
• Learning agent. It encodes five context dimensions and sends them into a dimension-aware mixture-of-experts (MoE). Each dimension has its own FFN and attention experts. A two-level gating scheme combines their outputs before an LLM predictor fine-tuned with LoRA.
• Companion agent. It turns expert activation weights into natural-language diagnostic reports, using an explicit base of rules drawn from educational theory.

According to the authors, the model outperforms mainstream KT baselines on real-world datasets. They also say it adapts well to context and produces understandable diagnoses. The work suggests that MoE routing weights can also serve as an explanation signal in educational AI.

Key points

• The five contexts are exercise difficulty, knowledge-concept relatedness, answer duration, error pattern and learning gain. Difficulty is computed from population correctness rates. Learning gain uses an Ebbinghaus-style retention decay. Response time is normalized with a default cap of 600.
• Gating works at two levels: one gate mixes FFN and attention experts within each dimension, and a sparse Top-K gate selects across dimensions. The authors contrast this design with RouterKT, which applies generic sequence MoE.
• The memory follows Tulving's split into episodic, semantic and procedural memory. A reflector component logs prediction deviations that exceed a threshold and issues targeted adjustments to gate parameters.
• The LLM components shown in the framework are GLM-4-flash for task parsing, reasoning and reflection, and a fine-tuned Llama-3-8B for response generation.
• Generated reports are rated on a 1/3/5 scale for diagnostic accuracy, comprehensibility and teaching feasibility. The extracted text does not include the dataset names or the numerical results.