Tobias Hey

Generative AI (GenAI) is rapidly transforming software engineering (SE) practices, influencing how SE processes are executed, as well as how software systems are developed, operated, and evolved. This paper applies design science research to build a roadmap for GenAI-augmented SE. The process consists of three cycles that incrementally integrate multiple sources of evidence, including collaborative discussions from the FSE 2025 “Software Engineering 2030” workshop, rapid literature reviews, and external feedback sessions involving peers. McLuhan’s tetrads were used as a conceptual instrument to systematically capture the transforming effects of GenAI on SE processes and software products. The resulting roadmap identifies four fundamental forms of GenAI augmentation in SE and systematically characterizes their related research challenges and opportunities. These insights are then consolidated into a set of future research directions. By grounding the roadmap in a rigorous multi-cycle process and cross-validating it among independent author teams and peers, the study provides a transparent and reproducible foundation for analyzing how GenAI affects SE processes, methods and tools, and for framing future research within this rapidly evolving area.

There are a multitude of software artifacts which need to be handled during the development and maintenance of a software system. These artifacts interrelate in multiple, complex ways. Therefore, many software engineering tasks are enabled — and even empowered — by a clear understanding of artifact interrelationships and also by the continued advancement of techniques for automated artifact linking. However, current approaches in automatic Traceability Link Recovery (TLR) target mostly the links between specific sets of artifacts, such as those between requirements and code. Fortunately, recent advancements in Large Language Models (LLMs) can enable TLR approaches to achieve broad applicability. Still, it is a nontrivial problem how to provide the LLMs with the specific information needed to perform TLR. In this paper, we present LiSSA, a framework that harnesses LLM performance and enhances them through Retrieval-Augmented Generation (RAG). We empirically evaluate LiSSA on three different TLR tasks, requirements to code, documentation to code, and architecture documentation to architecture models, and we compare our approach to state-of-the-art approaches. Our results show that the RAG-based approach can significantly outperform the state-of-the-art on the code-related tasks. However, further research is required to improve the performance of RAG-based approaches to be applicable in practice.

Traceability Link Recovery (TLR) is an enabler for various software engineering tasks. One important task is the recovery of trace links between Software Architecture Documentation (SAD) and source code. Here, the main challenge is the semantic gap between the two artifact types. Recent research has shown that this semantic gap can be bridged by using Software Architecture Models (SAMs) as intermediates. However, the creation of SAMs is a manual and time-consuming task. This paper investigates the use of Large Language Models (LLMs) to extract component names as simple SAMs for TLR based on SAD and source code. By doing so, we aim to bridge the semantic gap between SAD and source code without the need for manual SAM creation. We compare our approach to the state-of-the-art TLR approaches TransArC and ArDoCode. TransArC is the currently best-performing approach for TLR between SAD and source code, but it requires SAMs as an additional artifact. Our evaluation shows that our approach performs comparable to TransArC (weighted average F1 with GPT-4o: 0.86 vs. TransArC’s 0.87), while only needing the SAD and source code. Moreover, our approach significantly outperforms the best baseline that does not need SAMs (weighted average F1 with GPT-4o: 0.86 vs. ArDoCode’s 0.62). In summary, our approach shows that LLMs can be used to make TLR between SAD and source code more applicable by extracting component names and omitting the need for manually created SAMs.

Traceability information is a fundamental prerequisite for many essential software maintenance and evolution tasks, such as change impact and software reusability analyses. However, manually generating traceability information is costly and error-prone. Therefore, researchers have developed automated approaches that utilize textual similarities between artifacts to establish trace links. These approaches tend to achieve low precision at reasonable recall levels, as they are not able to bridge the semantic gap between high-level natural language requirements and code. We propose to overcome this limitation by leveraging fine-grained, method and sentence level, similarities between the artifacts for traceability link recovery. Our approach uses word embeddings and a Word Mover’s Distance-based similarity to bridge the semantic gap. The fine-grained similarities are aggregated according to the artifacts structure and participate in a majority vote to retrieve coarse-grained, requirement-to-class, trace links. In a comprehensive empirical evaluation, we show that our approach is able to outperform state-of-the-art unsupervised traceability link recovery approaches. Additionally, we illustrate the benefits of fine-grained structural analyses to word embedding-based trace link generation.

Classifying requirements is crucial for automatically handling natural language requirements. The performance of existing automatic classification approaches diminishes when applied to unseen projects because requirements usually vary in wording and style. The main problem is poor generalization. We propose NoRBERT that fine-tunes BERT, a language model that has proven useful for transfer learning. We apply our approach to different tasks in the domain of requirements classification. We achieve similar or better results F1-scores of up to 94%) on both seen and unseen projects for classifying functional and non-functional requirements on the PROMISE NFR dataset. NoRBERT outperforms recent approaches at classifying non-functional requirements subclasses. The most frequent classes are classified with an average F1-score of 87%. In an unseen project setup on a relabeled PROMISE NFR dataset, our approach achieves an improvement of 15 percentage points in average F1 score compared to recent approaches. Additionally, we propose to classify functional requirements according to the included concerns, i.e., function, data, and behavior. We labeled the functional requirements in the PROMISE NFR dataset and applied our approach. NoRBERT achieves an F1-score of up to 92%. Overall, NoRBERT improves requirements classification and can be applied to unseen projects with convincing results.