Pretrained models are usually adapted once, then frozen. I work on what comes after: adapting them to new tasks
and data without erasing what they already know. I mostly use continual learning, model merging and
parameter-efficient fine-tuning, on vision and vision-language models. More recently I've worked on data mixtures
for LLM pre-training.
PhD thesis · 2026
Beyond Fine-Tuning: Continual Adaptation of Pretrained Vision Models
Télécom Paris, Institut Polytechnique de Paris · Advised by Stéphane Lathuilière & Enzo Tartaglione
Papers
Imad Eddine Marouf, Khalid Oublal, Enzo Tartaglione, Stéphane Lathuilière
NeurIPS 2026 arXiv Code
Summary — DMC
Prompt learning pulls CLIP toward the classes it was trained on, while generalizing to new classes needs it to stay close to its zero-shot features. Both goals compete for the same prompt. DMC gives each goal its own prompt and merges them at test time, at the cost of a single prompt. Evaluated on 11 datasets.
Imad Eddine Marouf, Enzo Tartaglione, Stéphane Lathuilière, Joost van de Weijer
ICCV 2025 Paper arXiv Code
Summary — QUAD
QUAD tackles continual visual question answering by replaying only the questions from past tasks. No images are stored, which reduces memory and privacy costs. Evaluated on VQAv2 and NExT-QA.
Imad Eddine Marouf, Subhankar Roy, Stéphane Lathuilière, Enzo Tartaglione
CoLLAs 2025 arXiv Code
Summary — PLASTIC
PLASTIC uses test-time adaptation to help a model keep learning new classes after its first training session, without forgetting earlier ones.
Imad Eddine Marouf, Subhankar Roy, Enzo Tartaglione, Stéphane Lathuilière
ECCV 2024 · Oral Paper arXiv Code
Summary — CoFiMA
CoFiMA averages model weights across tasks, weighting each parameter by its Fisher information. A simple way to balance keeping old knowledge with learning new tasks.
Imad Eddine Marouf, Enzo Tartaglione, Stéphane Lathuilière
WACV 2024 Paper arXiv
Summary — MiMi
MiMi fine-tunes vision transformers with adapters that shrink step by step, guided by a neuron-importance score. Evaluated on 29 datasets.