BACE: Bloom’s Asymmetric Hybrid Cognitive-Level Classification Educational Large Language Model

Large language models (LLMs) are widely used in intelligent teaching and cognitive-level classification. However, mainstream Transformer architectures incur redundant attention computation, and a single feed-forward network struggles to perform the fine-grained routing of features for Bloom’s cognitive features. This study proposes BACE, an educational LLM for cognitive-level text classification based on an asymmetric hybrid architecture guided by Bloom’s taxonomy. Built on the pretrained ChatGLM3-6B, BACE replaces the MLP in the first 4 layers with a shallow Top-2 sparse mixture-of-experts (MoE) routing module to achieve token-level dispatching of cognitive features while retaining the self-attention mechanism. The middle 20 layers retain the native self-attention and MLP structures to maintain semantic transition. In the last 4 layers, attention is replaced with a deep, selective state-space modeling strategy, combined with chain-of-thought fine-tuning based on Bloom’s taxonomy, to integrate the global instructional context. Experiments on a public computer science course question dataset across multiple random seeds show that BACE achieves an average accuracy of 73.56% and a weighted F1 score of 0.6915, improving on ChatGLM3-6B (65.78%/0.6011) by 7.78 percentage points and 0.0904, respectively. Although its accuracy and weighted F1 are slightly lower than those of Qwen1.5-7B, BACE attains a BLEU-4 of 0.3426 and a ROUGE-L of 0.4601 on the generated chain-of-thought reasoning, both outperforming open-source models including Qwen1.5-7B, Mistral-7B, and Llama-3-8B. Through its asymmetric hybrid architecture, the model enables the classification of Bloom’s cognitive levels and provides a scalable technical pathway for question-answering-based cognitive-level assessment in intelligent education scenarios.