The AURA Research Lab — AI for Understanding, Reliable Automation — investigates how foundation models can be made measurably more efficient, interpretable, and trustworthy for software engineering. We work across model-centric and output-centric perspectives, from quantization and parameter-efficient fine-tuning to neurosymbolic reasoning and rigorous reliability evaluation.
AI systems that write our software should be cheap enough to deploy, honest about what they do not know, and accountable when they fail.
Foundation models are reshaping software engineering at every layer of the stack — from autocomplete to autonomous code review. But the same models are also expensive, opaque, and frequently wrong in ways that matter. Our lab is built to address that gap.
The work spans top-tier venues — ICSE, FSE, ASE, TSE, TOSEM, EMSE — and is anchored by an NSF CRII award and ongoing collaborations across William & Mary, USI Lugano, William & Mary's Charles Center program, and partner institutions.
AI for code · Understanding what models do · Reliable evaluation of their behavior · Automation that practitioners can trust. The threads cross-pollinate — efficiency feeds reliability, neurosymbolic reasoning feeds interpretability, and so on.
Quantization, parameter-efficient fine-tuning, and distillation that make code-intelligence models cheap enough to deploy at developer scale — without sacrificing functional or non-functional code quality.
Combining neural language models with grammars, type systems, and program analysis — plus feature-level explanations that move beyond opaque token-by-token predictions toward something developers can actually read.
Causal reasoning and counterfactual analysis to understand cause-and-effect in software systems — from flaky test debugging to LLM-as-a-Judge evaluation methodologies that measure what matters, not what's easy.
Designing autonomous agents that plan, reason, and execute multi-step software workflows — from issue triage to code review and test generation — with measurable end-to-end behavior on benchmarks like SWE-bench.
Funded research lines and ongoing student-led projects.
A two-year initiation award supporting research on making foundation models for code measurably more trustworthy and substantially less resource-intensive — through quantization, parameter-efficient fine-tuning, and rigorous reliability evaluation tailored to software engineering tasks.
Two summers of W&M Charles Center funding supporting undergraduate research on neural compression, LoRA-style fine-tuning, and code-quality evaluation under model-topology changes. Leon Huang's 2026 work also received a Monroe Scholars Conference Travel Grant.
A growing group of PhD students and undergraduates working at the intersection of AI and software engineering.
PhD from USI Lugano (2024). Research at the intersection of AI, NLP, and Software Engineering. Editorial board member of IEEE Computer, NSF CRII PI, and serves on 14+ program committees including ICSE, FSE, and ASE. Two ACM SIGSOFT Distinguished Reviewer awards.
