YMCQ: Reasoning-Enhanced MCQ Generation
Saved in:
| Title: | YMCQ: Reasoning-Enhanced MCQ Generation |
|---|---|
| Language: | English |
| Authors: | Andreea Dutulescu, Stefan Ruseti, Denis Iorga, Mihai Dascalu, Danielle S. McNamara (ORCID |
| Source: | Grantee Submission. 2025. |
| Peer Reviewed: | Y |
| Page Count: | 9 |
| Publication Date: | 2025 |
| Sponsoring Agency: | National Center for Education Research (NCER) (ED/IES) |
| Contract Number: | R305T240035 |
| Document Type: | Speeches/Meeting Papers Reports - Research |
| Descriptors: | Automation, Computer Assisted Testing, Multiple Choice Tests, Natural Language Processing, Test Items, Misconceptions, Efficiency, Open Source Technology, Test Validity |
| DOI: | 10.1007/978-3-031-98465-5_39 |
| Abstract: | Automated multiple-choice question (MCQ) generation is valuable for scalable assessment and enhanced learning experiences. How-ever, existing MCQ generation methods face challenges in ensuring plausible distractors and maintaining answer consistency. This paper intro-duces a method for MCQ generation that integrates reasoning-based explanations for both correct answers and distractors, leveraging open-source language models finetuned on publicly available datasets. Our approach addresses these issues with 3 major improvements. First, over 300k questions from public datasets were augmented with synthetically generated reasoning explanations. Second, we fine-tune a Large Language Model (LLM) with reasoning-based explanations to condition the generation while accounting for correct reasoning and possible misconceptions. Third, we introduce a multi-step filtering pipeline to ensure the validity of the question and the diversity of the generated distractors. This work argues for the effectiveness of reasoning-enhanced finetuning in improving MCQ generation quality while maintaining accessibility and cost efficiency. We release all resources, including synthetically augmented questions, training code, and the best model as open-source. [This paper was published in: "AIED 2025, LNAI 15882," edited by A. I. Cristea et al., Springer Nature Switzerland, 2025, pp. 308-15.] |
| Abstractor: | As Provided |
| IES Funded: | Yes |
| Entry Date: | 2025 |
| Accession Number: | ED674604 |
| Database: | ERIC |
| Abstract: | Automated multiple-choice question (MCQ) generation is valuable for scalable assessment and enhanced learning experiences. How-ever, existing MCQ generation methods face challenges in ensuring plausible distractors and maintaining answer consistency. This paper intro-duces a method for MCQ generation that integrates reasoning-based explanations for both correct answers and distractors, leveraging open-source language models finetuned on publicly available datasets. Our approach addresses these issues with 3 major improvements. First, over 300k questions from public datasets were augmented with synthetically generated reasoning explanations. Second, we fine-tune a Large Language Model (LLM) with reasoning-based explanations to condition the generation while accounting for correct reasoning and possible misconceptions. Third, we introduce a multi-step filtering pipeline to ensure the validity of the question and the diversity of the generated distractors. This work argues for the effectiveness of reasoning-enhanced finetuning in improving MCQ generation quality while maintaining accessibility and cost efficiency. We release all resources, including synthetically augmented questions, training code, and the best model as open-source. [This paper was published in: "AIED 2025, LNAI 15882," edited by A. I. Cristea et al., Springer Nature Switzerland, 2025, pp. 308-15.] |
|---|---|
| DOI: | 10.1007/978-3-031-98465-5_39 |
