MARQ unveils modular, LLM‑agnostic requirement quality assessor with systems‑engineering expertise
| Source: ArXiv | Original article
A new arXiv preprint introduces MARQ, a modular, LLM‑agnostic tool that assesses requirement quality using systems engineering knowledge.
A new arXiv pre‑print introduces MARQ, a modular, LLM‑agnostic framework designed to assess the quality of system‑engineering requirements. The authors argue that while large language models (LLMs) have shown promise for evaluating requirement texts, their off‑the‑shelf performance falls short of the rigor demanded by standards such as INCOSE. MARQ addresses this gap by layering a criterion‑specific knowledge stack on top of any LLM, thereby operationalising INCOSE quality criteria.
The prototype was tested on an expert‑labeled NASA requirement set, running 100 executions each with Claude Opus 5, GPT‑5.6 Sol and the open‑weight Gemma 3 12B model. Compared with baseline LLM outputs, MARQ consistently delivered higher alignment with the expert labels, demonstrating that domain‑specific engineering knowledge can materially boost LLM‑based assessment.
The development matters because requirement defects often cascade through the systems‑engineering lifecycle, inflating cost and risk. By improving automated quality checks, MARQ could reduce manual review effort, accelerate development timelines, and raise confidence in mission‑critical AI‑driven software—areas already highlighted in recent work on dynamic AI service‑chain optimisation.
Going forward, the research community will be watching for broader validation on diverse requirement corpora and for integration pathways with existing RE toolchains. Industry adoption could hinge on how easily the knowledge stack can be customised for different standards or project contexts, and whether the approach scales to larger, more complex specifications. If MARQ’s promise holds, it may become a cornerstone for embedding systems‑engineering discipline into the next generation of AI‑assisted development workflows.
Sources
Back to AIPULSEN