Requirements Elicitation in the Era of LLMs: A Survey on Large Language Models for Automated and Assisted Requirements Elicitation

Requirements Elicitation (RE) is critical but traditionally manual, costly, and prone to missing latent needs. This survey explores how Artificial Intelligence—particularly Large Language Models (LLMs) and Pre-trained Language Models (PLMs)—is transforming RE. We categorize recent approaches into Direct RE (AI actively elicits) and Indirect RE (AI supports training or mines existing data) by analyzing five distinct paradigms: (1) real-time, mistake-guided question generation for live interviews; (2) simulated user agents for uncovering latent needs; (3) privacy-preserving multi-agent fleets using small LLMs for goal-to-story decomposition; (4) generative script creation for RE education; and (5) lexicon-enhanced PLMs for mining noisy app reviews. Our analysis reveals that structured multi-agent systems enable small, private models to rival large LLMs, while domain-specific constraints (e.g., mistake-avoidance prompting, sentiment lexicons) significantly boost performance over generic fine-tuning. We highlight the trade-offs between generative flexibility and discriminative precision, outlining future directions for hybrid, privacy-aware AI in requirements engineering.