Machine learning researcher · London

Robert Hu

Portrait of Robert Hu

Research thesis

Efficient, grounded, interpretable learning.

Scaling is most useful when computation is efficient. I study stable numerics and hardware-aware algorithms for pre-training and inference. My longer-term goal is to bring these ideas together: scale models efficiently, then teach them to infer cause and effect rather than correlation alone—making their predictions better grounded and easier to interpret.

Current focus. Low-precision training and inference—particularly FP4—with an emphasis on stability and hardware efficiency.

Systems and numerics for scale

Numerical formats, GPU algorithms, and scalable methods that make training and inference more computationally efficient.

  1. 2026 · arXiv

    Fast Polynomial Transcendentals for LLMs

    Robert Hu

    Code fast-polynomial-transcendentals

  2. 2026 · arXiv

    Format-Aware Fusion for Fast FP4 Pretraining

    Robert Hu

    Code mfu-fp4-training

  3. 2026 · arXiv

    Hardware-Aware FP4 FlashAttention-4

    Robert Hu

    Code fp4-fa4

  4. 2026 · arXiv

    UE5M3 FP4 Block Scaling for Stable Language Model Pretraining

    Robert Hu, Carlo Luschi, Paul Balança

    Code ue5m3-fp4

  5. 2025 · arXiv

    Elucidating the Design Space of FP4 Training

    Robert Hu, Carlo Luschi, Paul Balança

  6. 2022 · NeurIPS

    Giga-scale Kernel Matrix-Vector Multiplication on GPU

    Robert Hu, Siu Lun Chau, Dino Sejdinovic, Joan Glaunès

    Code F3M

  7. 2022 · Machine Learning, 111

    Large Scale Tensor Regression Using Kernels and Variational Inference

    Robert Hu, Geoff K. Nicholls, Dino Sejdinovic

    Code KernelFriedTensor

Grounding and interpretation

Causal methods and interpretable representations for models that learn more than correlation and are easier to inspect.

  1. 2024 · Journal of Machine Learning Research, 25(160)

    A Kernel Test for Causal Association via Noise Contrastive Backdoor Adjustment

    Robert Hu, Dino Sejdinovic, Robin J. Evans

    Code kgformula

  2. 2024 · Transactions on Machine Learning Research

    Doubly Robust Kernel Statistics for Testing Distributional Treatment Effects

    Jake Fawkes, Robert Hu, Robin J. Evans, Dino Sejdinovic

    Code DR_distributional_test

  3. 2023 · Computational Mathematics and Data Science

    Towards Deep Interpretable Features

    Robert Hu, Dino Sejdinovic

    Code DIF

  4. 2022 · University of Oxford · DPhil thesis

    Large Scale Methods for Kernels, Causal Inference and Survival Modelling

    Robert Hu

    Code DIF · kgformula · KernelFriedTensor · Sumo-Net · F3M

  5. 2021 · Northern Lights Deep Learning Workshop

    Robust Deep Interpretable Features for Binary Image Classification

    Robert Hu, Dino Sejdinovic

    Code DIF-NLDL

* Equal contribution.

Background

Selected experience

Oxford profile ↗

2025 — 2026

Research Scientist

Graphcore, London

Research on low-precision pre-training, GPU kernels, and hardware-aware attention.

2022 — 2025

Applied Scientist

Amazon, London

Developed and deployed language-model and advertising-auction systems.

2022

Postdoctoral Research Assistant

University of Oxford

Worked on survival analysis for genomic data.

2018 — 2022

Industrial PhD

H&M and University of Oxford

Combined statistical research with large-scale forecasting and churn modelling.

2017 — 2018

Data Scientist

H&M, Stockholm

Built forecasting, experimentation, and decision-support systems.

2016 — 2017

Data Analyst

iZettle, Stockholm

Worked on data analysis and decision support.

Education

2014 — 2016

MSc in Mathematical Statistics

KTH Royal Institute of Technology

2014 — 2017

BSc in Business and Economics

Stockholm School of Economics

2011 — 2014

BSc in Engineering Physics

KTH Royal Institute of Technology

Contact