MARQ: A Modular, LLM-Agnostic Requirement Quality Assessor with Systems Engineering Knowledge Infusion Released

Summary:MARQ, a novel modular prototype, enhances LLM-based requirement quality evaluation by infusing Systems Engineering (SE) knowledge through a criterion-specific knowledge stack. Evaluated against off-the-shelf baselines using Claude Opus 5, GPT-5.6 Sol, and Gemma 3 12B on an expert-labeled NASA requirement set, MARQ achieves higher median F1 scores, particularly excelling in defect detection and false-alarm control for Correctness and Conformance criteria. However, the performance gains are model-

---

Technical Mechanism Analysis

The core mechanism of MARQ is to enhance LLMs with a criterion-specific knowledge stack that operationalizes Systems Engineering (SE) knowledge. This knowledge stack is designed based on INCOSE's requirement quality criteria and guides the reasoning process of LLMs to improve their performance in requirement quality assessment tasks. Specifically, MARQ achieves this through the following steps:

1. Knowledge Stack Design: The knowledge stack of MARQ includes specialized rules and logic for each quality criterion (e.g., Correctness, Completeness, Consistency), which are used to direct the reasoning process of LLMs.
2. Model Agnosticism: MARQ adopts a modular design, allowing it to be seamlessly integrated with various LLMs (e.g., Claude Opus 5, GPT-5.6 Sol, Gemma 3 12B).
3. Evaluation Process: During evaluation, MARQ first preprocesses the requirements using the knowledge stack, then the LLM performs reasoning, and finally, the knowledge stack's output is used to revise and optimize the LLM's reasoning results.

Engineering Trade-offs and Performance

MARQ significantly improves the performance of LLM-based requirement quality assessment, but it also faces some trade-offs and challenges:

1. Performance Improvement: Compared to off-the-shelf models, MARQ excels in defect detection and false-alarm control. For example, Claude Opus 5's F1 score increased from 0.55 to 0.80 after using MARQ.
2. Model Dependency: The performance improvement of MARQ is model-dependent. Claude Opus 5 shows excellent performance in false-alarm control, while GPT-5.6 Sol improves in detection but also increases false alarms. Gemma 3 12B's performance is close to that of off-the-shelf proprietary models but remains highly variable.
3. Computational Overhead: Injecting SE knowledge increases the computational overhead of the model, especially when processing large requirement sets, requiring more powerful computing resources.

Developer Implementation and Deployment Recommendations

For developers, MARQ provides a new approach to enhancing LLM performance in specific tasks by injecting domain knowledge. Here are some deployment recommendations:

1. Model Selection: Choose the appropriate LLM based on specific requirements. For example, if false-alarm control is a priority, Claude Opus 5 may be the preferred choice.
2. Knowledge Stack Customization: Customize the knowledge stack according to the actual application scenario to better adapt to the requirements of specific fields.
3. Resource Planning: Given the computational overhead of MARQ, it is recommended to plan sufficient computing resources, especially when processing large requirement sets.
4. Continuous Optimization: Regularly evaluate and optimize MARQ to adapt to changing domain requirements and technological developments.

Conclusion

MARQ demonstrates the potential of enhancing LLM performance in requirement quality assessment by injecting Systems Engineering knowledge. Although its performance improvement is model-dependent, it provides a new technical path for LLM applications in specific fields. In the future, with the continuous enrichment of SE knowledge and the development of LLM technology, MARQ is expected to play a greater role in more fields and scenarios.

---

Source: ArXiv AI (cs.AI) (2026-10-29)

— END —

Tags: #MARQ Follow #LLMs & Foundation Models Follow #Systems Engineering Follow #Requirement Quality Assessment Follow #Knowledge Infusion Follow