cv

PhD

Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

Computer Science

Applied Science Intern

Wayve

Worked on world modeling, VR-based driving-scene analysis, and 4D scene reconstruction. I developed a VR tool for interactively analyzing driving scenes and methods for reconstructing and iteratively refining dynamic driving scenes using Gaussian splatting.

MAD: Motion Appearance Decoupling for efficient Driving World Models

CVPR 2026

We propose a two-step approach to driving world models that separates generation into motion forecasting and appearance synthesis. We fine-tune off-the-shelf video generation models into two components: a motion forecaster that predicts future object movements in an abstract skeleton representation, and an appearance synthesizer that converts this motion into realistic video. Our approach is effective and efficient, requiring only 140 GPU hours and driving videos for training, being two orders of magnitude less than prior work.

From Generation to Generalization: Emergent Few-Shot Learning in Video Diffusion Models

ICML 2026

We introduce a few-shot fine-tuning framework that repurposes pretrained video diffusion models for novel visual and logic tasks using only a handful of examples. The updated version of this work, Rethinking Visual Intelligence: Insights from Video Pretraining, received an ARC Prize 2025 Honorable Mention.

GEM: A Generalizable Ego-vision Multimodal World Model for Fine-Grained Ego-Motion, Object Dynamics, and Scene Composition Control

CVPR 2025

We propose a generalizable ego-vision multimodal world model, GEM, that predicts future frames of egocentric driving videos using multiple control signals. We build on a pretrained Stable Video Diffusion model, adding new components for fine-grained controllability and long-term generation.

A Multi-Loss Strategy for Vehicle Trajectory Prediction: Combining Off-Road, Diversity, and Directional Consistency Losses

Under review

Traditionally, trajectory prediction models are trained using minADE losses, which only penalize the closest prediction to the ground truth. This sparsification of the loss function slows down model convergence and may lead to suboptimal predictions for less trained prediction heads. In this work, we propose a multi-loss strategy that combines off-road, diversity, and directional consistency losses which act on all prediction modes and improve the performance of trajectory prediction models.

Sim-to-Real Causal Transfer: A Metric Learning Approach to Causally-Aware Interaction Representations

CVPR 2025

We investigate causal understanding in the context of multi agent interaction prediction, where removing non-causal agents from the scene should not change model's prediction. We show modern prediction models are able to identify non-causal agents in the scene, but fail to properly model causal agent removals. We propose a metric learning approach to learn causally-aware interaction representations, which we show lead to better generalizability and robustness for Out of Distribution (OOD) and low-data regime scenarios.

Vehicle trajectory prediction works, but not everywhere

CVPR 2022

We show that current trajectory prediction models fail to fully understand scene structure, where naturalistic perturbations in the scene, like introducing turns, can significantly affect the prediction quality of the model. We propose a scene attack method to evaluate the robustness of trajectory prediction models against such perturbations, and further show improvements of the models when trained with these adversarial examples.

ARC Prize 2025 Honorable Mention

ARC Prize Foundation

Recognized for research on repurposing video diffusion models for novel visual and logic tasks.

Third Place in LLM Hackathon at EPFL

EPFL

A two-day hackathon on improving LLM architectures like GPT-2 and Llama 2 held at EPFL.

Bronze Medal

In the National Olympiad in Informatics.

Ranked Second in !Optimizer Contest

A nation-wide scientific contest in optimization.

Head

Cultural Student Association, EPFL

Led cultural and scientific events with 250+ participants.

Co-organizer

Mathematical Sciences Summer School

Organized a summer school for 80 high-school students covering mathematics and computer science topics.

Head

Student Scientific Association

Led student scientific activities in the Department of Mathematical Sciences.