Daniele Potertì

I’m a final-year PhD candidate in Artificial Intelligence at the University of Milano-Bicocca, supervised by Fabio Mercorio. I’m currently a visiting PhD student at NTU Singapore in Erik Cambria’s group.

I work on understanding and evaluating large language models through their internals:

• Steering and role vectors: activation directions that shape model behaviour (EMNLP Findings 2025)
• Evaluating interpretability methods: do automatic explanations of sparse autoencoder features match what the features actually do? (EMNLP 2025)
• Deception and safety: a theory-driven approach to studying deceptive behaviour in LLMs (in progress), and LLM fragility and malicious outputs (EMNLP 2026)
• LLM evaluation and benchmarks, especially for Italian (ITALIC, NAACL 2025)

Before my PhD I spent four years as a freelance software developer (Go, Java, C++, Python).