Advances in Educational Large Language Models for Youth Programming and Personalized Learning

Dr. Su Yu, senior engineer and associate researcher at the Hefei Institute of Artificial Intelligence, introduced the background and challenges of intelligent education, focusing on youth programming, data sparsity, and learning latency.

The talk highlighted three technical innovations: (1) training a youth‑programming domain large language model using dual data and historical experience injection; (2) enabling small‑knowledge learning through a layered knowledge graph and prompt generation; (3) applying reinforcement‑based cognitive recommendation by simulating learning environments with large models.

Product cases were presented, including the Frog Programming Platform, an AI programming learning machine, and a digital‑human AI recorded‑course platform, all integrating the aforementioned technologies.

Technical details covered data acquisition (generating realistic error code via adversarial networks), model fine‑tuning on LLaMA with LoRA, knowledge injection using historical repair cases stored in an embedding vector library, and evaluation interfaces that score code similarity and test‑case passing.

The session also discussed the limitations of large models—high computational resources, long training time, and difficulty of customization—and proposed strategies for smaller companies, such as model pruning, quantization, knowledge distillation, and leveraging domain‑specific expertise and fine‑grained data to build competitive vertical LLMs.

Finally, a Q&A addressed training and fine‑tuning methods for educational LLMs and how to augment sparse user knowledge using adversarial neural networks to simulate additional learning data.