About

Hi, I'm Nastya. Currently I'm a machine learning research engineer at Apple Machine Learning Research , working on MLX and efficient training and inference of LLMs.

I am interested low precision training and quantization, GPU kernel programming, distributed training and inference, and local AI. In the past I did research in AI for neuroscience at the Mathis Laboratory of Adaptive Intelligence and trained models for wearable devices at the Machine Intelligence Lab.

Background

I completed my MSc in Data Science at EPFL. While there I interned at Apple Machine Learning Research in California, under Samy Bengio, working with Ronan Collobert and his team on efficient training and inference of large language models.

I was also a research assistant at the Mathis Laboratory of Adaptive Intelligence, led by Prof. Mackenzie Mathis, working on self-supervised pre-training on neural and behavioural data with Steffen Schneider, and on foundation models for animal pose estimation with Shaokai Ye. Before that I was a quantitative research intern at WorldQuant in New York, fine-tuning LLMs on proprietary financial data.

I completed my BSc in Applied Mathematics and Physics at the Moscow Institute of Physics and Technology, where I worked at the Machine Intelligence Lab on training and deploying models on wearable devices — phones, fitness watches, trackers.

Publications

1. Optimal Splitting of Language Models from Mixtures to Specialized DomainsSkyler Seto, Pierre Ablin, Anastasiia Filippova, Jiayuan Ye, Louis Béthune, Angelos Katharopoulos, David GrangierICML 2026
2. Stochastic KV Routing: Enabling Adaptive Depth-wise Cache SharingAnastasiia Filippova, David Grangier, Marco Cuturi, João MonteiroICLR 2026 Workshop on Test-Time Updates
3. Scaling Categorical Flow MapsOscar Davis, Anastasiia Filippova, Pierre Ablin, Victor Turrisi, Amitis Shidani, Marco Cuturi, Louis BéthunearXiv preprint, 2026
4. Partial Parameter Updates for Efficient Distributed TrainingAnastasiia Filippova, Angelos Katharopoulos, David Grangier, Ronan CollobertNeurIPS 2025 Workshop on Optimization for Machine Learning
5. No Need to Talk: Asynchronous Mixture of Language ModelsAnastasiia Filippova, Angelos Katharopoulos, David Grangier, Ronan CollobertICLR 2025 (Spotlight, 3.2%)
6. Time-series attribution maps with regularized contrastive learningSteffen Schneider, Rodrigo González Laiz, Anastasiia Filippova, Markus Frey, Mackenzie W. MathisAISTATS 2025
7. SuperAnimal pretrained pose estimation models for behavioral analysisShaokai Ye, Anastasiia Filippova, Maxime Vidal, Steffen Schneider, Jessy Lauer, Tian Qiu, Alexander Mathis, Mackenzie Weygandt MathisNature Communications, 2024

Full list on Google Scholar.

Other projects

• Neuro Sense — activity recognition and heart-rate estimation on a tiny AI chipActivity representations and heart-rate estimation from PPG and IMU data on a smartwatch, with the activity encoder trained by self-supervised contrastive learning. 5 bpm mean absolute error, 4% MAPE — and at 5M MAC operations the chip running it draws 0.1 mW.

Elsewhere

• GitHub
• Scholar
• Twitter
• LinkedIn
• Email
• CV