Hey, I'm Giorgos Nikolaou !

MSc in Data Science, EPFL • Researcher at ISTA

About Me

I'm currently living in Vienna 🇦🇹, working as a Research Intern at ISTA with Francesco Locatello while pursuing my MSc in Data Science at EPFL, after completing my BSc in Computer Science at the University of Athens. When I'm not buried in assignments or research, you can find me swimming, experimenting with new recipes, or catching up on some much-needed sleep. My research focuses on understanding the internal representations and behaviour of frontier models and using these insights to inform approaches in AI safety.

Representation Learning CoT Faithfulness AI Safety

Experience

Research Intern

Institute of Science and Technology Austria (ISTA)

PI: Francesco Locatello

Sep 2026 - Present
  • Developing a formal account of why chain-of-thought (CoT) reasoning can be unfaithful.
  • Investigating methods to improve CoT faithfulness and characterizing their inherent tradeoffs.

Research Student

Theory of Machine Learning Lab (TML), EPFL

PI: Nicolas Flammarion

Feb 2026 - Jul 2026
  • Developed a theoretically grounded, unsupervised method for selective LLM unlearning (NeurIPS 2026).
  • Extended open-unlearning with token-importance-weighted losses and additional benchmarks (code).

Research Assistant

GLADIA, Sapienza Università di Roma

PI: Emanuele Rodolà

Aug 2025 - Jan 2026
  • Proved language models are almost surely injective and used this insight for exact model inversion (ICLR 2026; tweeprint).
  • Research focused on understanding and operationalizing the representation space of modern LLMs.

Data Science Intern

Logmind

Jul 2025 - Jan 2026
  • Optimized and scaled the log-based anomaly detection pipeline to support 200x data volume without degrading performance.
  • Prototyped an LLM-driven system for automated log analysis, generating actionable insights to streamline incident response.

Research Assistant

Archimedes Research Unit

PI: Yannis Panagakis

Jan 2025 - Mar 2025
  • Studied image-to-image translation, leveraging multi-attribute embeddings to condition generation.
  • Applied multi-modal learning techniques to enhance performance and flexibility of generative models.

Data Engineering Intern

INNOV-ACTS

Mar 2024 - Aug 2024
  • Built optimized data-processing pipelines to extract, transform, and validate multi-organization data for the EuroHyPerCon project.
  • Developed automation scripts that streamlined data-integration workflows and delivered clear, actionable analytical outputs.

Education

Master of Science in Data Science

École Polytechnique Fédérale de Lausanne (EPFL)

Sep 2024 - Present
  • Maintaining a strong GPA of 5.77/6.
  • Completed advanced coursework in Measure-Theoretic Probability, Statistics, Optimization Theory, and Deep & Reinforcement Learning.

Bachelor of Science in Computer Science

National and Kapodistrian University of Athens

Oct 2020 - Jul 2024
  • Graduated with distinction, earning a GPA of 9.18/10.
  • Thesis selected among the department's best Bachelor's and Master's theses (2024 cohort).
  • Thesis: "Higher-Order Deep Unfolding Networks for Compressed Sensing"
    Supervisor: Yannis Panagakis
    Developed a deep unfolding network for compressed sensing leveraging proximal algorithms within the Augmented Lagrangian framework, integrating a novel higher-order polynomial module to enhance model expressiveness. Achieved state-of-the-art reconstruction performance with over 3dB PSNR improvement at ultra-low sampling rates (1-10%).

Awards & Grants

  • Research Grant ($80,000)

    Coefficient Giving — Administered by Berkeley Existential Risk Initiative.

    2026
  • Postgraduate Scholarship

    Bodossaki Foundation

    2025
  • Outstanding Bachelor's and Master's Thesis Distinction

    University of Athens — 2024 cohort.

    2025
  • Postgraduate Scholarship

    Union of Greek Shipowners (UGS)

    2024

News

Oct 02, 2026
Preprint Alert: The Independence Prior of SAEs Fragments Visual Concepts. We show that standard SAEs implicitly treat concepts across image patches as independent, and introduce Spatial-SAE, which models the missing spatial dependencies to recover more coherent and interpretable visual concepts; read it on arXiv!
Sep 24, 2026
Our paper "Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance" was accepted at NeurIPS 2026, see you at Sydney!
Jul 01, 2026
Our paper "Language Models are Injective and Hence Invertible" was selected for an oral presentation at the Greeks in AI 2026 Symposium, held at the Eugenides Foundation in Athens (July 15–17).
Jun 04, 2026
Preprint Alert: Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance. We introduce a theoretically grounded, unsupervised method to identify which tokens matter for forgetting and use it to drive more selective LLM unlearning. Read it on arXiv.
Jan 26, 2026
Our paper "Language Models are Injective and Hence Invertible" was accepted at ICLR 2026!
Nov 27, 2025
Invited talk on LLM injectivity and invertibility at the Laboratory of Data Engineering of Area Science Park, where we presented and discussed our recent work. Many thanks to Alberto Cazzaniga and Diego Doimo for the invitation and warm hospitality!
Oct 17, 2025
Preprint Alert: Language Models Are Injective and Hence Invertible. We prove that LLMs are injective, empirically stress-test this property, and develop the first inversion algorithm with theoretical guarantees. To top it off, the announcement went viral on (>5M views)!
Jul 07, 2025
Excited to begin my Data Science Internship at Logmind!
Jun 27, 2025
Honored to have received the Bodossaki Foundation Graduate Scholarship to support my graduate studies at EPFL.
May 25, 2025
My undergraduate thesis was selected for publication among the department's best Bachelor's and Master's theses for 2024!
Jul 04, 2024
Grateful to have been awarded the "SYN-ENOSIS for Education Scholarship" by the Union of Greek Shipowners, supporting my graduate studies at EPFL.