TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring

AI Digest

该研究提出TACT框架，通过分类法对LLM进行微调，提升英语教学的适应性。结合双分类体系构建数据集，优化模型策略，显著优于现有基准。

This paper introduces TACT, a framework using taxonomies to enhance LLMs for adaptive ESL tutoring. It constructs annotated datasets and optimizes models through taxonomy-aligned training, outperforming existing benchmarks.

Key points

• 构建双分类体系（教学策略/学生行为）增强模型适应性 Develops dual taxonomies for tutor strategies and student behaviors
• 开发TACTCorpus数据集包含3.2万条标注对话数据 Constructs TACTCorpus with 32,379 annotated dialogues
• 采用分组相对策略优化提升教学支架质量 Applies group-relative policy optimization for scaffolding
• 在78个教学场景中实现20.3%性能提升 Achieves 20.3% performance gain in 78 tutoring scenarios

Takeaway: 分类法指导的微调显著提升LLM英语教学的适应性 / Taxonomy-guided fine-tuning significantly enhances LLMs' adaptability in ESL tutoring

Why it matters 为教育AI提供可复用的分类框架与评估标准，具有实际应用价值