I study how neural networks model the physical world, and how they learn together.

I am a researcher at Bielefeld University, working with Barbara Hammer; my PhD is funded by the Honda Research Institute Europe. After a Visiting Fellowship in Philip Torr’s Torr Vision Group at the University of Oxford, I continue to work with the group on world models for video. Recently, GeoPhys took first place on both Physics‑IQ video-to-video leaderboards, and two of my papers were accepted at NeurIPS 2026. I review for NeurIPS, ICML and IJCNN, and have mentored master’s students at the Honda Research Institute.

My research focuses on world models and physical reasoning: I read the geometry of learned representations to find out what neural networks know about physics, and use it to tell whether a generated video could really happen. Broadly, I want physical plausibility to become something we can measure: principled ways to read physics out of a model without changing it, and to put that structure to work as a prior for learning, a verifier at inference time and a diagnostic for anomalies.

Alongside this, I work on distributed learning, where many models learn together without sharing their data; that line includes an oral at the NeurIPS 2024 workshop on federated foundation models. I have been a visiting researcher at Cold Spring Harbor Laboratory and Leiden University, and I studied data science at the University of Milano-Bicocca. Off the clock, you’ll find me surfing, at the cinema or at a concert.

Recently

  1. Dec 2026

    Next: NeurIPS 2026 in Sydney.

  2. Oct 2026

    Is a Linear Probe Evidence of a Linear Representation? is on arXiv.

  3. Oct 2026

    Released the GeoPhys feature dataset on Hugging Face.

  4. Sep 2026

    Two papers accepted at NeurIPS 2026: GeoPhys, and the linear probe paper as joint first author.

  5. Sep 2026

    Two new preprints: Procedural Core and MotionJEPA.

  6. Jun 2026

    GeoPhys reached first place in video-to-video generation on both Physics‑IQ leaderboards.

Earlier news
  1. Jun 2026

    Started a Visiting Fellowship in Philip Torr’s group at the University of Oxford.

  2. Jun 2026

    Geodesic-SVDD Latent Mixup for One-Shot Distributed Tabular Learning at IJCNN 2026 in Maastricht.

  3. Feb 2026

    The Observer Effect in World Models is on arXiv.

  4. Jan 2026

    PPFLex at the AAAI 2026 workshop on federated learning for critical applications.

  5. Dec 2025

    At NeurIPS 2025 in San Diego with ReStraV, and the Narcissus Hypothesis at a workshop.

  6. Dec 2024

    Oral at the NeurIPS 2024 International Workshop on Federated Foundation Models.

Selected publications

World models and physical reasoning

  1. NeurIPS 2026

    GeoPhys: The Geometry of Physical Plausibility

    C. Internò, A. Pondaven, H. Issa, F. Pizzati, F. Pinto, M. Olhofer, I. Laptev, P. Torr, E. P. Simoncelli, B. Hammer, D. Klindt

    arXivcodedataproject page

    First on both Physics‑IQ video‑to‑video leaderboards

  2. NeurIPS 2026

    Is a Linear Probe Evidence of a Linear Representation?

    N. Kutluozen*, C. Internò*, D. Klindt

    arXiv
  3. NeurIPS 2025

    AI-Generated Video Detection via Perceptual Straightening

    C. Internò, R. Geirhos, M. Olhofer, S. Liu, B. Hammer, D. Klindt

    arXivcode

    Cited in a forensic report by Hany Farid

Distributed learning

  1. IoT-J 2026

    Pruning Federated Models Through Loss Landscape Analysis and Client Agreement Scoring

    C. Internò, E. Raponi, M. Olhofer, A. Raza, T. Bäck, N. van Stein, Y. Jin, B. Hammer

    IEEE Internet of Things JournaljournalarXivcode

    Oral at the NeurIPS 2024 workshop FL@FM

  2. IJCNN 2024

    Federated Loss Exploration for Improved Convergence on Non-IID Data

    C. Internò, M. Olhofer, Y. Jin, B. Hammer

    paperarXivcode

All publications and preprints