LLM-Driven Grading System for K-12 Non-Native English Learners

A new framework adapts large language model outputs to the proficiency levels of K-12 English learners in non-native contexts, using China's national curriculum (CSE) as a case study. The system controls lexical complexity via a four-tier grading system, supported by graded vocabulary lists and a multi-turn dialogue corpus. The core technical contribution is the DDPO algorithm (Diversity Driven Policy Optimization), a multi-turn GRPO-based approach that preserves dialogue diversity while optimizing quality. DDPO achieves low out-of-vocabulary rates and high diversity, enhancing conversational naturalness and pedagogical effectiveness. The research addresses the widespread challenge of proficiency mismatch in LLMs for education.