I’m an AI Scientist Intern at Mistral AI, working on multimodal large language models across pretraining and post-training. I’m also a PhD candidate at the University of Amsterdam / ELLIS, advised by Yuki Asano and Cees Snoek.
My research focuses on vision–language learning, video understanding, and computational efficiency.
Previously, I was a Student Researcher at Google DeepMind and a Research Intern at Amazon Web Services (AWS) on the Rekognition team.
First-author and co-first-author papers. * denotes equal contribution.
We introduce PIN, a learned spatial prompt that enables frozen vision–language models to localize objects. Training uses synthetic data and next-token prediction, without manually annotated detection data.
Adaptation of OpenFlamingo and BLIP-2 · Zero-shot object localization
We introduce TVBench, a 10-task benchmark for temporal understanding in video–language models. It addresses shortcuts in existing benchmarks that let models answer from text or a single frame.
We introduce SnapViT, a structured pruning method that adapts pretrained vision transformers to different compute budgets without retraining or labeled data.
Evaluated on DINO, SigLIPv2, DeiT, and AugReg models.
We jointly learn a video encoder and its prediction targets through masked modeling. Balanced clustering prevents representation collapse and provides targets that capture semantic and temporal information.
arXiv 2025
Steering frozen language models with a single intervention in the key–value cache, without fine-tuning or changing the prompt.
TMLR 2025
Using LLMs to optimize prompts for downstream vision–language tasks.
CVPR Workshops 2022
CVPR 2021
CVPR 2021
ICCV 2021
CVPR 2021
Nature Machine Intelligence 2021
CVPR 2020
Research across industry and academia, with a foundation in representation learning and computational physics.
2026–present
AI Scientist Intern
Research on pretraining and post-training multimodal large language models.
2026
Student Researcher
Research on efficient video tokenization and pretraining.
2022–present
PhD Candidate · QUVA Lab / ELLIS
Spatial grounding in vision–language models, temporal evaluation, self-supervised video representations, and model efficiency. Supervised by Yuki Asano and Cees Snoek.
2021
Research Intern · Rekognition
Self-supervised video representation learning. First-author work on shuffled contrastive learning (SCVRL).
2019–2022
MSc Computational Physics · CompVis
Generative modeling with Björn Ommer’s group, including image-to-video synthesis with conditional normalizing flows. Research visit with Kosta Derpanis in Toronto.
For research inquiries and collaborations: