DLPP: a generative learning path planning framework based on LLM semantic guidance and closed-loop optimization

Abstract

Generating personalized learning paths remains an open problem in intelligent education systems, where conventional sequence recommendation methods typically output ranked resource ID lists without encoding pedagogical intent or supporting post-deployment refinement. To address these limitations, we present DLPP (Diffusion-based Learning Path Planning Framework), which recasts path generation as a two-stage process of instructional planning followed by semantic instantiation. In the first stage, a conditional diffusion model produces structured activity-type sequences within a continuous embedding space; nearest-neighbor quantization then maps these sequences to discrete pedagogical categories. In the second stage, a Large Language Model (LLM) equipped with Retrieval-Augmented Generation (RAG) converts each abstract plan step into a concrete, resource-grounded learning activity. An ensemble of five student behavior simulators supplies uncertainty-aware quality scores, and a Diffuser-based reinforcement learning module closes the optimization loop. Evaluated on two public educational datasets, DLPP yields 7–8% PKG improvement over the pure diffusion baseline on both EdNet and Junyi Academy. On the medium-scale EdNet (2941 training paths), it outperforms RL-Path and DiffPath baselines while remaining competitive with rule-based methods.