Integrating Prediction and Generation: A Lightweight Large Language Model Framework With Parameter Fusion for Personalized Feedback
TL;DR
A lightweight LLM-based framework designed for joint performance prediction and feedback generation and a Low-Rank Adaptation-based parameter-efficient fine-tuning mechanism is proposed, indicating that parameter fusion and parameter-efficient fine-tuning provide effective solutions for integrating prediction and feedback generation in educational environments.
Abstract
Predicting student performance and generating personalized pedagogical feedback are critical yet computationally demanding tasks in educational artificial intelligence (AI). Although large language models (LLMs) excel at both numerical prediction and text generation, their deployment is often hindered by high inference costs and difficulty in integrating heterogeneous learning objectives. This study proposes a lightweight LLM-based framework designed for joint performance prediction and feedback generation. The framework consists of three key components: 1) a feature engineering scheme that maps 11 categories of Massive Open Online Course (MOOC) learning behaviors into structured semantic representations, 2) a post-training parameter fusion strategy that decouples task-specific optimization before parameter fusion, and 3) a Low-Rank Adaptation (LoRA)-based parameter-efficient fine-tuning mechanism. We evaluated the framework using a real-world dataset containing 1,527 students from a Java programming MOOC. The experimental results demonstrate that behavioral feature-guided prompting enables high-quality feedback generation (ROUGE-1 F<inline-formula> <tex-math notation="LaTeX">$1=0.76$ </tex-math></inline-formula>). For student performance prediction, the standalone prediction model achieved an R2 value of 0.9882. Compared with direct joint multi-task learning, the proposed parameter fusion strategy maintained a strong predictive performance (R<inline-formula> <tex-math notation="LaTeX">${}^{2} =0.9624$ </tex-math></inline-formula>) while preserving feedback generation quality. Furthermore, LoRA reduced the number of trainable parameters by more than 99.9%. The LoRA-based adaptation substantially reduced total training time while maintaining robust cross-cohort generalization performance (R<inline-formula> <tex-math notation="LaTeX">${}^{2} =0.9064$ </tex-math></inline-formula>). These results indicate that parameter fusion and parameter-efficient fine-tuning provide effective solutions for integrating prediction and feedback generation in educational environments.