Decoupled Mode Connectivity for Base-to-Novel Generalization in Vision-Language Models
NeurIPS 2026
Learns two prompts, one for seen classes and one for new ones, and merges them at test time for the cost of a single prompt.
I'm an AI researcher working on continual learning: how pretrained models can keep adapting to new tasks without forgetting what they already know. I recently completed my PhD at Télécom Paris, advised by Stéphane Lathuilière and Enzo Tartaglione. During the PhD I was a student researcher on Google DeepMind's Gemma team, working on data mixtures for pre-training, and a visiting researcher at the Computer Vision Center in Barcelona, hosted by Joost van de Weijer.
I also write about research and what I learn along the way, including from building CurriboxAI. I'm open to research roles on LLM and vision-language model training: get in touch.
NeurIPS 2026
Learns two prompts, one for seen classes and one for new ones, and merges them at test time for the cost of a single prompt.
ICCV 2025
Continual visual question answering that replays only past questions, without storing any images.
CoLLAs 2025
Uses test-time adaptation so a model keeps learning new classes after its first training session.
ECCV 2024 · Oral
Averages model weights across tasks, weighting each parameter by its Fisher information.
WACV 2024
Fine-tunes vision transformers with adapters that shrink during training, guided by neuron importance.