Nhi Pham

I'm a PhD student at the Max Planck Institute for Informatics in Saarbrücken, Germany, where I am advised by Prof. Bernt Schiele and Dr. Jonas Fischer. As part of PhD preparatory phase, I was jointly supervised by Dr. Adam Kortylewski on interpreting image classifiers with 3D neural object volumes.

Before this, I received my B.S in Computer Science and Mathematics at New York University, where I worked on computational linguistics with Prof. Adam Meyers, and on operator theory with Prof. Ilya Spitkovsky. I also interned at Goldman Sachs, Amazon AWS AI, and Meta (Facebook).

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

News

09.2026  Great turnout at our ECCV 2026 workshop World Models in the Loop - thanks to our speakers and attendees!

05.2026  Our paper CAVE received an oral presentation ✨ at XAI4CV@CVPR 2026! See you in Denver!

04.2026  Our workshop on World Models in the Loop is accepted at ECCV 2026. See you in Malmö!

04.2026  Accepted at ICVSS 2026. See you in Sicily!

01.2026  Our paper CAVE is accepted at ICLR 2026. Check it out!

04.2025  I presented CAVE at the Generative Intelligence Lab.

03.2025  I’ve finished my PhD preparatory phase in the IMPRS-TRUST program, and officially joined D2 MPII for my PhD.

12.2024  H-POPE is accepted to NeurIPS 2024 - Statistical Foundations of LLMs and Foundation Models Workshop!

Featured Research

My current research interest lies in geometric representation learning, with an emphasis on interpretable and robust vision models grounded in 3D structure. I am also interested in extending these ideas to video diffusion models, where learning holistic, geometry-aware tokens can support structured reconstruction and generation over time. In my previous life, I enjoyed playing around with special matrices and their eigenvalues.

Design an inherently-interpretable and robust classifier by extending existing 3D-aware classifiers with concepts extracted from their volumetric representations for classification. We also propose a new metric to measure 3D consistency of concepts.

A coarse-to-fine-grained benchmark that systematically assesses hallucination in object existence and attributes.

The known constructive tests for the shapes of the numerical ranges in the 3-by-3 case are further specified when the matrices in question are row stochastic. Auxiliary results on the unitary (ir)reducibility of such matrices are also obtained.

Talks

05.2026  Efficient Visual Latent Representation, Computer Vision and Multimodal Learning Un-Workshop, Tübingen AI Center

04.2025  Escaping Plato's Cave: Seeing through Explicit Representation, Generative Intelligence Lab

Others

Source code and design are borrowed from Jon Barron's website.