Can AI Build Systems? an Exploratory Study on Generating Software Architecture With LLMS
Bibliographic record
Abstract
Software architecture is instrumental in shaping the quality, maintainability, and integrity of a software system, providing a blueprint that ensures that all components function cohesively. Despite being introduced only recently, Large Language Models (LLMs) have shown great potential in automating low-level software engineering tasks, such as coding and test generation. However, the application of LLMs on high level cognitive tasks, such as architecture and design, has not been explored as much. To that end, this study evaluates the capability of LLMs-specifically Llama 4 Llama, Gemini 2.0 Gemini, and OpenAI o3 in generating software architecture given unstructured textual requirements in natural language. We employ prompt engineering techniques, including zero-shot prompting, in-context learning and chain-of-thought prompting to assess each model's responsiveness and accuracy. Furthermore, we evaluate iterative refinement, an approach where an LLM revises the output of another for enhanced quality, akin to a human editing a preliminary draft for improvements. Our results indicate that state of the art models like OpenAI o3 and Gemini 2.0 Gemini perform exceptionally well in identifying all the domain concepts of the system and how they interact, especially in conjunction with iterative refinement, while Llama' Llama 4's performance was subpar. In terms of generating an architecture for the system, again models from OpenAI and Gemini are able to come close to the benchmark architecture by using the correct architectural styles especially with Chain-of-Thought prompting, however Llama's model do not infer the correct styles.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.006 | 0.038 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.005 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".