FACULTY
APPLY NOW
GRAD
STUDENTS
APPLY
BY DEC 15
AI & SE @ NC State
AI, carefully: AI enthusiasm plus engineering rigor. That is how a top-five software engineering group uses AI. Reading our 2025–26 papers, five themes emerge. One warning, before the hype: this faculty recently co-edited an IEEE Software special issue on AI and productivity whose bottom line was “AI coding tools are genuinely useful — but they do not substitute for human judgment.” That attitude — that we should carefully assess both the benefits and costs of AI — runs through all our work.
1. Humans and AI, side by side
We model how real developers actually work with AI — and how AI could join the team. Kuttal’s HFXE lab predicts where collaborating developers will navigate next (CHI 2026 Honorable Mention) and maps how LLM early adopters are reshaping software work (EMSE 2026). Stolee and Xu jointly modeled 76 developers’ intentions, actions, and emotions while coding with AI (FSE 2026), and Xu’s human-AI code comprehension work took a Distinguished Paper Award at ICPC 2026.
2. Testing the AI
AI is software; somebody has to check it. D’Amorim tests the deep learning libraries everything else depends on, via fuzzing and neurosymbolic constraint learning (ICSE 2026). Ore measures how dependably LLMs generate code. Xu probes privacy leaks and backdoors in code LLMs (IEEE TCSE Distinguished Paper, ICSME 2026). Assunção checks whether LLMs can even tell when two programs are the same.
3. AI as the engineer
We put LLM agents to work on the chores developers dread: end-to-end dependency upgrades (ICSME 2026), agentic runtime healing (CACM 2026), and LLM-assisted backporting for the Robot Operating System (Assunção + Xu, ICSE 2026) — always with a human in the loop.
4. Frugal AI: how low can you go?
Against the bigger-is-better tide: Menzies’ data-light challenge shows a few dozen labels often match massive optimizers across 100+ SE tasks; the EZR toolkit packs a dozen classic learners into under 1000 lines; the MOOT repository gives everyone the data to check us. Kuttal’s lab made the same point from the HCI side: a sub-10ms behavioral model beat a 32B-parameter LLM at predicting developer navigation (VL/HCC 2026).
5. Teaching SE in the AI era
Our stance: we do not need classes on how to prompt; we need classes on how to govern. In the AI era, our graduates need to:
- understand code more deeply, and how it can be improved;
- treat testing as imperative when using AI tools;
- understand both symbolic and subsymbolic (LLM) systems;
- know when an LLM service can be replaced by something cheaper;
- integrate AI tools into existing systems.
Heckman’s six-year study of a CS2 course traces how students’ help-seeking moved from forums and office hours to GenAI (SIGCSE 2026), and Stolee and Ore study how students comprehend and review code — the exact skill that matters most when the code was written by a machine.