Giovanni Pinna

AI Researcher & Engineer | Ph.D. in Applied Data Science & AI

University of Trieste

Welcome!

I am Giovanni Pinna, an AI Researcher and Engineer based in Trieste, Italy. I hold a Ph.D. in Applied Data Science & Artificial Intelligence from the University of Trieste (March 2026).

My research focuses on the intersection of Natural Language Processing, Large Language Models, and Evolutionary Computation — particularly on improving LLM-generated code through Genetic Improvement techniques and developing evaluation metrics for Text-to-SQL systems.

I have international research experience at University College London (UCL) in London, UK and NOVA IMS in Lisbon, Portugal. I am the author of 10+ publications in top international venues including Scientific Reports (Nature), IEEE Access, and EuroGP.

🗞️ News

• Apr 2026 Published two papers “Comparing ai coding agents: A task-stratified analysis of pull request acceptance” and “Analyzing Message-Code Inconsistency in AI Coding Agent-Authored Pull Requests” at MSR 2026 .
• Mar 2026 Completed my Ph.D. in Applied Data Science & AI at the University of Trieste!
• Sep 2025 Started visiting research at University College London (UCL), in Prof. Federica Sarro's group.
• 2025 Published “Redefining Text-to-SQL Metrics” in Scientific Reports (Nature) and 2 papers at SSBSE 2025 (GA4GC and HotCat).

🎓 Ph.D. Thesis

Application of Large Language Models: Addressing Real-World Challenges

Ph.D. Thesis — University of Trieste, 2025

Large Language Models promise to reshape how we write code, query data, and access knowledge — but do they really deliver? This thesis probes that question across four fronts: how we evaluate LLMs, how they democratize expertise, how we make them reliable, and how we deploy them sustainably. A human-centered study on legal texts shows that domain experts still outperform state-of-the-art models precisely where the stakes are highest. To close the reliability gap, I introduce a Genetic Improvement framework that systematically repairs LLM-generated code, and a continuous Text-to-SQL metric that uncovers distinctions hidden by pass-or-fail scores. On sustainability, multi-objective optimization discovers coding-agent configurations with over a hundredfold hypervolume improvement. The result: a path to LLM deployment that is not only powerful, but trustworthy. Supervisors: Prof. Luca Manzoni, Prof. Andrea De Lorenzo.

📄 Selected Publications

• Redefining Text-to-SQL Metrics by Incorporating Semantic and Structural Similarity
  G. Pinna, Y. Perezhohin, L. Manzoni, M. Castelli, A. De Lorenzo
  Scientific Reports 15.1 (Nature), 2025
  Paper
• Comparing AI Coding Agents: A Task-Stratified Analysis of Pull Request Acceptance
  G. Pinna, J. Gong, D. Williams, F. Sarro
  arXiv:2602.08915, 2026
  arXiv
• Analyzing Message-Code Inconsistency in AI Coding Agent-Authored Pull Requests
  J. Gong, G. Pinna, Y. Bian, J. M. Zhang
  arXiv:2601.04886, 2026
  🏆 Distinguished Mining Challenge Paper Award, MSR 2026
  Paper
• Enhancing Large Language Models-Based Code Generation by Leveraging Genetic Improvement
  G. Pinna, D. Ravalico, L. Rovito, L. Manzoni, A. De Lorenzo
  EuroGP 2024, Springer LNCS vol. 14631
  Paper
• An Artificial Intelligence System for Automatic Recognition of Punches in Fourteenth-Century Panel Painting
  M. Zullich, V. Macovaz, G. Pinna, F.A. Pellegrino
  IEEE Access, 2023
  Paper

📝 Recent Posts

🚀 Projects

🏷️ Popular Topics

🛠️ Skills