Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Large language models (LLMs) can match or even outperform human agile experts in writing fluent, well-formatted user stories from app reviews when using few-shot prompts.result 90%
• LLMs struggle to produce independent and unique user stories from app reviews, which are essential for building a strong agile backlog.limitation 85%
• Using few-shot prompting (one-shot and two-shot) significantly improves LLM adherence to the canonical user story template 'As a [role], I want [feature] so that [benefit]' compared to zero-shot prompting for models like GPT-3.5 Turbo.technical 88%
• The smaller model Mistral 7B Instruct performs best in adhering to the canonical user story template under zero-shot conditions, and its performance degrades when few-shot examples are provided.result 82%
• LLM-generated user stories from app reviews can achieve high scores on automated quality criteria like atomicity, minimality, and conceptual soundness, often matching or exceeding scores associated with human-written stories.result 87%
• Gemini 2.0 Flash with two-shot prompting generates the highest quality user stories from app reviews according to human evaluation using the RUST framework.result 86%
• LLMs can reliably turn unstructured, messy app reviews into actionable software requirements in the form of user stories, providing developers with clear guidance for improvement.result 90%
• Existing automated techniques for analyzing app reviews often do not perform well when replicated and fail to produce clean, backlog-ready user stories for agile projects.limitation 80%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card