Pietro Bernardelle

PhD Candidate, The University of Queensland

pb_pic.jpg

Room 622 - Building 78

The University of Queensland

Brisbane, Australia

I’m a second-year Computer Science PhD candidate at The University of Queensland, supervised by Professor Gianluca Demartini, Dr. Joel Mackenzie, and Dr. Kevin Roitero.

My research broadly focuses on understanding bias and information representation in large language models (LLMs), with particular interests in political and perspective bias, relevance assessment, multilinguality, and model internals. Outside of my specific research interests, I’m mostly just very curious and enjoy learning about things I know little about. If you have an interesting idea, project, or problem you’d like to chat about, I’m always happy to hear about it — and if you happen to run into me at a conference, definitely come say hi!

My recent works and collaborations include:

  • Investigating how LLMs internally encode query–document relevance and whether these representations transfer across languages (CIKM 2026).
  • Studying whether LLMs develop shared, language-independent representations of problem difficulty (ACL 2026).
  • Developing a framework for measuring LLMs perspective alignment effectiveness across downstream applications (NLPerspectives@LREC 2026).
  • Influencing political ideology in LLMs using synthetic personas (The Web Conference 2025).

selected publications

  1. Political Ideology Shifts in Large Language Models
    Pietro Bernardelle, Stefano Civelli, Leon Fröhling, Riccardo Lunardi, and 2 more authors
    In Information Processing & Management (IPM), Elsevier, 2026
    arXiv
  2. LLMs Encode Relevance as a Layer-Wise Cross-Lingual Signal
    Pietro Bernardelle, Samaneh Mohtadi, Stefano Civelli, Joel Mackenzie, and 1 more author
    In Proceedings of the 35th International ACM Conference on Knowledge and Information Management (CIKM), 2026
    arXiv
  3. Persona Conditioning as an Assessor-Sensitivity Probe for LLM-Based IR Evaluation
    Samaneh Mohtadi, Pietro Bernardelle, Joel Mackenzie, and Gianluca Demartini
    In Proceedings of the 35th International ACM Conference on Knowledge and Information Management (CIKM), 2026
    arXiv
  4. A Shared Geometry of Difficulty in Multilingual Language Models
    Stefano Civelli, Pietro Bernardelle, Nicolò Brunello, and Gianluca Demartini
    In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Jul 2026
    DOI arXiv Bib