ExpertGenQA: Open-ended QA Generation in Specialized Domains

[Download Paper]

Generating expert-quality questions in narrow technical domains is harder than it looks: few-shot prompting is either shallow or slow, and LLM judges tend to reward fluent writing over content. ExpertGenQA combines few-shot examples with topic and question-style categories so generated QA pairs stay diverse and cognitively closer to expert writing. On U.S. Federal Railroad Administration documents it doubles the efficiency of standard few-shot prompting at 94.4% topic coverage, and retrieval models trained on the resulting queries improve top-1 accuracy by 13 points.