Location: Télécom Paris, Institut Polytechnique de Paris (France)
Host Laboratory: S2A team / LTCI lab
Supervising Team: Quentin Bouniot (Main Supervisor, Télécom Paris), Prof. Florence d’Alché-Buc (Thesis Director, Télécom Paris), Prof. Dr. Stefanie Jegelka (TUM)
Starting Date: Fall 2026 (Flexible)
Duration: 3 years

Topic

Understanding and Characterizing the Geometry of Representations Across Layers, Models, and Modalities

Foundation Models (large vision or language models, multimodal architectures, etc.) have become central to modern AI, but their internal process of data remain poorly understood. This PhD project aims to develop a principled framework for analyzing the internal representations of these models, layer by layer, across architectures, and across data modalities. The goal is to build theoretically grounded tools that characterize how these models process information, and to establish actionable links between internal structure and downstream properties such as transferability, robustness, and interpretability.

More details on the research topic can be found here.

What We Are Looking For

We are seeking a highly motivated candidate with:

Required:

• A Master’s degree (or equivalent) in machine learning, applied mathematics, computer science, or a related field
• Strong foundations in mathematics (linear algebra, probability, optimization, statistics)
• Solid programming skills (Python, PyTorch or similar deep learning framework)
• Interest in deep learning theory, representation learning, or interpretability
• Good communication skills in English (written and spoken)

Optional:

• Previous experience in research in deep learning (e.g., publications)
• Experience working with Foundation Models, Representation-level analysis or on mechanistic interpretability

What We Offer

• A vibrant research environment at Télécom Paris, one of France’s top engineering schools, within the Institut Polytechnique de Paris and French research ecosystem
• An international collaboration with the Technical University of Munich (TUM), with opportunities for research visits in Munich
• Funding for conference attendance and publications at top-tier venues (NeurIPS, ICML, ICLR, CVPR, etc.)

How to Apply

Please send the following documents to quentin[dot]bouniot[at]telecom-paris[dot]fr:

• CV (including a list of relevant coursework and any publications)
• Transcripts (Master’s and Bachelor’s)
• A brief motivation letter explaining your interest in the topic (1 page max)
• Contact information for 1–2 references

Applications will be reviewed on a rolling basis. Do not hesitate to reach out informally if you have questions about the position or the research topic.

We are committed to fostering an inclusive research environment and encourage applications from candidates of all backgrounds.