Artificial Intelligence and Software Engineering
Reviewed: 2025-10-10
SERL advances two complementary directions in the areas of artificial intelligence (AI) and software engineering (SE): using AI to make software work smoothly, and engineering AI-intensive systems to be dependable in the real world. On the “AI for SE” side, the group shows how NLP and LLMs help teams classify and understand requirements at scale, including multi-label classification across large taxonomies and automatic detection of causal relations in requirements to support impact reasoning. They frame what “good” requirements look like with a model of fitness-for-purpose and quality attributes to guide everyday decisions.
Similarly, for testing and maintenance, we use LLMs to speed up the labeling of GUI test cases and make web-element localization more robust to change, while multi-agent setups explore which LLMs best fit GUI-test generation. Our work on bug-report triage uses ML to spot invalid bug reports early in the maintenance cycle. We also explore what it takes to adopt ML solutions in practice. We have also used AI beyond code generation, e.g., we use AI to help prioritize code reviews and detect industry-relevant code smells in large codebases.
On the “SE for AI” side, the group identifies practical risks and proposes safeguards in the development, maintenance and operations of AI based solutions. We have proposed the use of “data smells” to expose data quality pitfalls, MLOps pipelines for continuous AI delivery, and explored security/assurance angles such as threat modeling and adversarial ML in industry. Like the “AI for SE” theme, we also study real adoption - how practitioners use AI. In this regard, we have focused on how GenAI helps in startups and product development. We have also investigated pain points across ML-enabled engineering to focus on improvements.
Current and Future Work
Future research directions include enabling tighter, safer pipelines for AI features in products - linking data-quality checks and threat models directly into CI/CD. We expect more task-specific LLM assistants in requirements, testing, and defect management, with evidence-based selection of models and prompts. We are also interested in developing practical quality metrics for GenAI systems and creating patterns for facilitating the adoption of AI in industry contexts.
Overall, our research helps teams ship AI-enabled software and traditional software systems that are faster to build, easier to test, and safer to operate, utilizing AI - grounded in what actually works in industrial settings.