As software systems continue to grow in size and complexity, especially in areas such as cloud computing, edge computing and the Internet of Things, there is an increasing need for adaptation at runtime. Although dynamic reconfiguration of software architectures aims to support structural changes, many existing approaches rely on centralised control, which can limit scalability and introduce single points of failure. In this sense, decentralised reconfiguration becomes appealing. This PhD work defines a set of research questions that guide the formalisation, design and evaluation of a decentralised approach to dynamic reconfiguration of compositional software architectures. This work started from a systematic literature review, which revealed some limitations in current approaches, including the need to improve reconfiguration time while ensuring behavioural correctness and non-functional optimality w.r.t. some formal specification. In particular, this research proposes a novel framework that introduces the notion of software architecture generators as a formal way to represent alternative single-head-single-tail software architectures and hierarchical control structure. The latter is intended to choose an optimal architecture in a decentralised manner.
Ensuring compliance with the European Union Artificial Intelligence Act (AI Act) presents a novel challenge for software architects. The Act introduces legally binding obligations for high-risk AI systems that must be addressed during system design, yet its legal language is often too generic and disconnected from software engineering practices. As a result, architects face difficulty in translating legal mandates into concrete architectural decisions, risking late-stage redesigns or non-compliance. In this work, we discuss the need for a decision-support tool that helps architects integrate the requirements of the AI Act during the early design phase of AI systems. We first isolate legal obligations that impact quality attributes, and then we contextualize those obligations to software architecture tactics and relevant industrial standards. We discuss the design and implementation of a web-based prototype that operationalizes this mapping, guiding architects through system classification under the AI Act and offering tailored architectural recommendations. A preliminary evaluation of a prototype provided encouraging feedback on the tool’s relevance, clarity, usability, and coverage. This work is a first step towards a comprehensive architecture design advisory tool, which helps software architects and engineers comply with AI Act recommendations. We conclude the paper by discussing future work and the potential of Generative AI to enhance the tool’s capabilities.
Large language models are increasingly applied to software architecture-related activities, such as architecture design and recovery. However, academia has not yet systematically classified how LLMs are employed in such architecture-specific activities. Hence, a clear and comprehensive understanding of the state of the art remains limited and fragmented. This article aims to classify studies on the application of LLMs in software architecture. We conducted a systematic mapping study based on established guidelines to answer three research questions. After careful filtering, 24 primary studies (out of 1,507) were selected from three scientific databases. Our main results indicate that most studies focus on modeling and maintenance activities, primarily leveraging source code and documentation as architectural inputs. These studies focus on conceptual approaches and empirical analyses, often relying on case studies and validation experiments conducted in controlled settings. However, critical phases such as requirements analysis and operations are largely underexplored. Furthermore, there is a notable lack of architecture-specific datasets, standardized evaluation frameworks, and robust tool support, limiting the reproducibility and practical applicability of current approaches and empirical studies. We note that no consolidated benchmark or widely adopted evaluation method was proposed, revealing a gap in the field. Our findings highlight limitations on LLM adoption, which can be explored by the research community. Finally, this article proposes a classification map and outlines key challenges and future research directions to support the development of empirically grounded AI-driven architectural solutions.
Software architects must make critical design decisions early in development, yet identifying which requirements actually drive those decisions remains a manual, error-prone and time-consuming process. Large software requirement documents contain hundreds of statements, yet only a fraction meaningfully shape system architecture and distinguishing these from the rest demands significant expert effort that does not scale to modern and large software projects. Existing automated approaches treat architecturally significant requirements (ASR) identification and architectural tactic recommendation as separate and disconnected activities, leaving architects without complete automated support at precisely the point where it is needed most. No established pipeline exists that moves from raw natural language requirements directly to justified and ranked architectural tactic recommendations. We propose a four-stage methodology pipeline that addresses this gap. Building on a transformer-based ASR classifier trained on curated 10,489 expert-annotated requirement artifacts, the pipeline maps each identified ASR to its implied quality attribute and queries an indexed SEI architectural tactic catalog using retrieval-augmented generation (RAG). Llama 3 then ranks and justifies retrieved tactics in the context of the specific requirement. We evaluate the pipeline on 312 held-out ASR instances drawn from 20 withheld software requirements specification documents. Comparing RAG pipeline recommendations against expert-validated labels using precision, recall, F1-score, Mean Reciprocal Rank (MRR) and Recall@3, the RAG pipeline achieves F1 = 0.78, MRR = 0.81 and Recall@3 = 0.91, outperforming a zero-shot Llama 3 baseline (F1 = 0.61) by 17 points. An ablation study confirms that quality attribute filtering contributes 7 F1 points of the overall gain. This work represents an initial step toward supporting architectural decision making through evidence-based architectural tactic recommendations offering a scalable and evidence-based approach to early-phase architectural decision support.
Detecting design patterns in source code is essential for understanding and maintaining software architectures, yet remains difficult due to variability in real-world implementations. This paper introduces a neuro-symbolic framework for design pattern detection that combines interpretable structural analysis with learned semantic representations. The approach integrates two complementary signals, i.e., topological conformance via abstract syntax tree analysis under a four-pillar heuristic, and semantic similarity to canonical pattern prototypes in a discriminative latent space. This latent space is obtained by fine-tuning CodeBERT with supervised contrastive learning, yielding cohesive and separable pattern clusters. To mitigate data scarcity and lexical bias, we generate and validate a synthetic dataset using large language models and apply identifier anonymization at the abstract syntax tree level. Evaluation on an unseen synthetic test set and on a manually validated Gold Standard benchmark demonstrates improved robustness over purely structural methods, with strong performance across standard metrics. Overall, combining structural constraints with optimised semantic representations offers a robust and scalable solution for design pattern detection in software architecture analysis.
