A Multi-Agent Framework for Feature-Constrained Difficulty Control in Reading Comprehension Item Generation

In three linesMAFIG, a multi-agent framework, uses multiple LLM agents and feature-specific evaluators to generate reading comprehension items with robust difficulty control. The method constructs sequences of feature constraints yielding monotonically increasing difficulty, outperforming existing single-agent approaches.

Summary generated by Claude — human-verified