Can LLMs Model Incorrect Student Reasoning? A Case Study on Distractor Generation
Researchers investigated whether LLMs can mimic human mistakes by generating plausible multiple-choice distractors. Testing models like DeepSeek-V3 and GLM-4 on the Eedi math dataset, the study found they use a 'solve-first' strategy. Providing the correct answer in prompts improved performance by 8%, helping models anchor their error simulations.