Luiza Corpaci
semantic faithfulness · formal methods · AI safety
I study whether meaning survives translation: when intent becomes a spec, a spec becomes code or a proof, or a model’s output becomes an explanation, does the content still mean what it meant, or does it just look like it does?
Current work:
- Research mentoring at MARS V and SPS Apart; participating in CAISH’s Hardware Assurance Program.
- Semantic faithfulness across representations: semantic diffing via the Equational Theories Project, measuring whether a representation preserves its underlying meaning across changes in formality.
- Other active threads: semantic equivalence in LLM-generated formal code, feature stability at rule-gap boundaries, bias geometry in embedding space.
Background: hardware verification (AMD), chip design (Marvell), combinatorial testing (SBA Research, Vienna), mechanistic interpretability (University of Vienna). Best Paper at ICTSS 2024. Completing an MSc in Data Science at TU Wien.
Availability: open to research roles and a PhD start from autumn 2027. Applications are in progress this cycle, so conversations about funding, supervision, and introductions are useful to me now rather than later.
Contact
Write to me at luiza.corpaci@gmail.com. I read everything, and I answer questions about the work, counterexamples, and collaboration first.
I’m open to fellowship and collaboration conversations around semantic faithfulness, formal methods, and AI safety. You can also find me on GitHub, LinkedIn, and Google Scholar.
AI hardware verification engineer
Wanderer by night.
Wonderer in the beyond.
Tarski, Gödel, and Chaitin are recurring reference points in my conceptual writing.
I wonder about wonders.
A question I keep nearby: whether the capacity to wonder tells us anything about the limits of formalization.