Date
Publisher
arXiv
With the rapid development of large language models (LLMs), the applications of LLMs have grown substantially. In the education domain, LLMs demonstrate significant potential, particularly in automatic text generation, which enables the creation of intelligent and adaptive learning content. This paper proposes a new LLMs framework, which is named as Reading Comprehension Exercise Generation (RCEG). It can generate high-quality and personalized English reading comprehension exercises automatically. Firstly, RCEG uses fine-tuned LLMs to generate content candidates. Then, it uses a discriminator to select the best candidate. Finally, the quality of the generated content has been improved greatly. To evaluate the performance of RCEG, a dedicated dataset for English reading comprehension is constructed to perform the experiments, and comprehensive evaluation metrics are used to analyze the experimental results. These metrics include content diversity, factual accuracy, linguistic toxicity, and pedagogical alignment. Experimental results show that RCEG significantly improves the relevance and cognitive appropriateness of the generated exercises.
Study at a glance
- Question: Can a fine-tuned model pipeline with post-processing effectively generate high-quality, controlled English reading comprehension exercises?
- Setting and data: 24,636 RACE questions, 437 Chinese Gaokao and Postgraduate exam passages (2010–present), 2,000 generated samples, and reasoning benchmark datasets.
- Method: Model training (supervised fine-tuning, reward modeling, PPO) with attribute-guided decoding and toxicity filtering, evaluated with automated metrics.
Why it matters
While large language models are increasingly used for teaching assistance, their use in automatically generating assessment content remains limited by poor task adaptability and insufficient control over output quality.
Key findings
- Fine-tuned models paired with reinforcement learning consistently improved text similarity scores, with the 3B model achieving a BLEU-4 of 27.23 compared to 1.39 for a fine-tuned GPT-2 baseline.
- Trained models achieved notable zero-shot reasoning improvements on HotpotQA and HellaSwag, such as a 1.5B model improving from 27.02 to 41.39 on HellaSwag.
- Post-processing guidance and toxicity filtering consistently lowered perplexity across clean and toxic test sets, dropping perplexity from 37.76 to 29.59 on highly toxic inputs for a 1.5B model.
Summary drafted with AI from the full paper and may contain errors. See the full paper for details.
What is the application?
Who age?
Why use AI?
Study design
