Automated generation of Software Requirements Specifications (SRS) is a field that has gained significant relevance with the emergence of large language models (LLMs). However, evaluating the quality of these generated specifications remains challenging, mainly because few datasets offer detailed traceability between source requirements, intermediate artifacts, and final specifications. In this context, ReqGenX emerges, an empirical study that proposes transforming legacy SRS documents into traceable synthetic pre-SRS artifacts. This approach enables a more granular evaluation of LLM-based SRS generation, exposing the inevitable trade-offs between faithfulness, information retention, and artifact completeness.
The ReqGenX methodology involves decomposing SRS sections into atomic statements, each linked to its original source. These statements are classified into artifact types inspired by standards through a plurality voting system using multiple LLMs. Subsequently, artifacts are generated using constrained prompts and iterative judge-guided refinement. The results of the study, applied to seven SRS documents from the PURE dataset, show AlignScore values between 0.96 and 0.99 for atomic statements, and between 0.80 and 0.94 for generated artifacts. Judge pass rates exceed 90% in most cases, although stricter evaluations with Prometheus reveal pass rates ranging from 54.8% to 97.1%.
What is interesting about this work is that it not only measures artifact quality but also evaluates their recoverability in a downstream SRS reconstruction. In a case study, artifact-backed atoms remained recoverable, with SBERT means between 0.69 and 0.75 and AlignScore medians between 0.76 and 0.84. This demonstrates that traceability can be maintained even when generating complete specifications from these artifacts.
For a software development company like Q2BSTUDIO, these findings have direct implications. The ability to decompose legacy requirements into traceable atomic components allows improving the quality of custom application projects. By integrating artificial intelligence techniques into the elicitation and specification process, we can offer our clients more precise solutions aligned with their real needs.
Atomic decomposition not only facilitates evaluation but also improves change management in complex projects. When a requirement changes, it is possible to identify exactly which atomic statements are affected and, by extension, which artifacts and specifications need updating. This is fundamental in custom application development, where requirements constantly evolve. At Q2BSTUDIO, we apply this principle in our requirements engineering processes, using traceability tools that benefit from atomic granularity.
Another relevant point is the ability of LLMs to generate artifacts from atomic statements. ReqGenX demonstrates that, with iterative refinement, high-quality artifacts can be obtained. However, the study also reveals an inverse relationship between faithfulness to the source and artifact completeness. In a business context, this trade-off must be carefully managed. For example, when generating specifications for a safety-critical system, faithfulness is paramount, while in rapid prototyping, completeness may be prioritized. Our teams at Q2BSTUDIO are trained to balance these criteria according to client needs.
Cloud infrastructure is a key enabler for implementing the ReqGenX approach at scale. LLMs require significant computational resources, and platforms like AWS and Azure offer elastic environments that adjust to demand. Additionally, the ability to run multiple model instances in parallel speeds up the plurality voting and refinement process. At Q2BSTUDIO, we design cloud architectures that allow our clients to leverage these capabilities without worrying about server management.
In the field of cybersecurity, atomic traceability allows precise security impact analysis. Each statement can be evaluated against security policies and regulations. If an artifact contains sensitive information or potential vulnerabilities, it can be traced back to its origin and patches or mitigations applied. Our pentesting service integrates with these practices, ensuring that generated specifications are secure by design.
Business intelligence also benefits. Traceability data can be visualized in Power BI dashboards, showing the status of each requirement, its origin, and associated artifacts. This provides a complete project view to stakeholders, improving decision-making. At Q2BSTUDIO, we implement BI solutions that connect directly with requirements management tools, offering real-time reports.
Finally, the incorporation of AI agents represents the next frontier. Imagine agents that not only generate artifacts but also negotiate with stakeholders to clarify ambiguous requirements, or automatically verify consistency among atomic statements. Although ReqGenX uses an iterative judge, the next evolution could include autonomous agents that make refinement decisions. At Q2BSTUDIO, we are researching how to integrate these agents into our development workflows, using our artificial intelligence services to increase productivity and reduce errors.
In summary, the ReqGenX study offers a roadmap for improving automated requirements generation through atomic decomposition and traceability. For development companies like Q2BSTUDIO, this methodology aligns perfectly with our focus on quality, transparency, and efficiency. We invite our clients to explore how we can apply these techniques in their custom software, cloud, cybersecurity, BI, and AI projects.
