Internship Research / World Models for Networking

Assignment
World Models are a family of models that offer an alternative to classical approaches such as vision models or large language models. Instead of mapping inputs to outputs, they learn an internal representation of a “world” and predict how that world evolves over time and in response to actions. This makes them natural candidates for prediction, simulation and planning.
These models are known to be hard to train. Many recent approaches build on the Joint-Embedding Predictive Architecture (JEPA) [1–4], which predicts in a learned representation space rather than reconstructing raw inputs. Within this line of work, LeWorldModel (LeWM) [5, 6] was recently proposed. It trains end to end in a stable way and has shown good properties for training.
Objectives
The LeWM methodology is general, but so far it has mainly been applied to image inputs. This internship will explore how it can be adapted to networking problems. Wireless networks are a good fit for this approach: their state, such as topology, connectivity and traffic, changes over time and in response to decisions. Prior work in the team has studied related problems [7].
Work plan
Study the World Model and JEPA literature, then reproduce and test the existing LeWM model.
Choose a suitable networking application and an input representation, based on feasibility and relevance.
Adapt and retrain the model on this application, possibly through a sequence of tasks of increasing complexity.
Evaluate the results experimentally and analyze both the predictions and the learned representations.
Candidate profile
Solid background in machine learning and deep learning, with practical experience in Python and PyTorch
Interest in networking (wireless networks are a plus, but not required)
Familiarity with self-supervised learning, representation learning or reinforcement learning is a plus
References
[1] Y. LeCun, “A Path Towards Autonomous Machine Intelligence,” OpenReview, 2022.
[2] M. Assran et al., “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture” (I-JEPA), CVPR 2023.
[3] A. Bardes et al., “Revisiting Feature Prediction for Learning Visual Representations from Video” (V-JEPA), TMLR 2024.
[4] M. Assran et al., “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning,” arXiv:2506.09985, 2025.
[5] R. Balestriero and Y. LeCun, “LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics,” arXiv:2511.08544, 2025.
[6] L. Maes, Q. Le Lidec, D. Scieur, Y. LeCun, and R. Balestriero, “LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels,” arXiv:2603.19312, Mar. 2026.
[7] F. Marcoccia, V. Fagoo, G. M. de Saint Julien, C. Adjih, T. Watteyne, and P. Mühlethaler, “NetDiff: Graph Diffusion with Improved Global Capabilities to Generate and Update Mobile Network Topologies,” ICML 2026.