Efficient deep learning PhD student @ 馃幁 GLADIA Research Group
My works:
-
Zero-Shot Quantization via Weight-Space Arithmetic 馃搫 Paper 路 馃捇 Code 路 馃 Checkpoints
Robustness to post-training quantization is a transferable direction in weight space and can be shared across datasets.
NeurIPS 2026 -
Weight Pair Encoding: Inducing a Smaller Grammar in Neural Network Weights 馃搫 Paper
Neural network weights can be perturbed to result in smaller, variable-length grammars across off-the-shelf compressors.
Preprint, 2026 -
The Undetected Damage of Quantization on Retrieval and How to Fix It 馃搫 Paper
A recipe to turn unreliable outputs of quantized models into reliable answers using mixed precision in classification and retrieval tasks. Preprint, 2026 -
On Task Vectors and Gradients 馃搫 Paper 路 馃捇 Code
Task vectors are a close approximation of multi-task learning gradients.
UniReps Workshop (PMLR, vol. 322) -
ATM: Improving Model Merging by Alternating Tuning and Merging 馃搫 Paper 路 馃捇 Code
Alternating finetuning and merging improves the merged models when using task vectors
ICIAP 2025 Workshops