Conjecture and Inquiry: Quantifying Software Performance Requirements via Interactive Retrieval-Augmented Preference Elicitation
New IRAP method turns vague requirements into math in 5 rounds
Researchers Wang Shi Hai and Chen Tao tackle a persistent software engineering challenge: translating vague, natural-language performance requirements into precise mathematical functions. Their new system, IRAP (Interactive Retrieval-Augmented Preference Elicitation), uses retrieval-augmented generation to iteratively query stakeholders, reducing ambiguity while minimizing cognitive load. The approach explicitly derives problem-specific knowledge to guide reasoning and progressive interactions, ensuring requirements are quantified accurately without overwhelming users.
Tested against 10 state-of-the-art methods across 4 real-world datasets, IRAP delivers up to 40x performance improvements in as few as 5 interaction rounds. Accepted at ACL 2026, this work bridges a critical gap in automated software engineering by converting fuzzy human preferences into actionable, quantifiable specifications. For tech teams, this means faster, more reliable performance requirement analysis with less back-and-forth.
- IRAP quantifies performance requirements into mathematical functions using interactive retrieval-augmented preference elicitation
- Achieves up to 40x improvements over 10 state-of-the-art methods on 4 real-world datasets
- Requires as few as 5 interaction rounds to reduce cognitive overhead and ambiguity
Why It Matters
Automates a tedious manual step in software engineering, cutting time and errors in requirement quantification.