Post-training & adaptation
Knowledge distillation, parameter-efficient fine-tuning, and memory-efficient learning for VLMs and LLMs.
Roy Miles
Senior Research Scientist @ Huawei Noah's Ark Lab UK · London, United Kingdom
Making post-training for vision-language and language models more memory-efficient and practical. I am interested in how distillation, parameter-efficient fine-tuning, and new training strategies can lower the cost of adapting capable models without compromising quality. The goal is to turn those advances into reliable methods that work under real product and hardware constraints.
Current focus: efficient foundation models
Multimodal learning
Visual representation learning, visual grounding, and visual-language model pre-training.
Mobile & on-device inference
Compact, fast vision models designed for resource-constrained devices and product cameras.
Currently exploring test-time scaling for diffusion language models.
Education & Experience
Senior Research Scientist — Huawei Noah's Ark Lab UK (2023 – Present)
Post-training research for VLMs and LLMs, covering distillation, PEFT, and memory-efficient training. Developed product-adopted fine-tuning methods for large-scale vision models.
Research Scientist — Samsung Research UK (2022 – 2023)
Developed semi-supervised video object segmentation for mobile devices. The work was integrated into a Samsung mobile product and received the SRUK 2023 Best Paper Award.
PhD, Computer Vision — Imperial College London (2018 – 2023)
Thesis: Towards Resource Efficient Vision Models. Supervised by Prof. Krystian Mikolajczyk.
MEng, Electrical & Electronic Engineering — University of Bristol (2014 – 2018)
First Class Honours. Thesis on an FPGA-based pipeline for object classification using binarized neural networks.
Selected Papers
Do You See What I Am Pointing At? Gesture-Based Egocentric Video Question Answering
CVPR 2026 — Gesture-based egocentric video question answering
Test-Time Scaling with Diffusion Language Models via Reward-Guided Stitching
arXiv preprint, 2026 — Diffusion language models and test-time scaling
Region-based Cluster Discrimination for Visual Representation Learning
ICCV 2025 — Highlight
VeLoRA: Memory Efficient Training using Rank-1 Sub-Token Projections
NeurIPS 2024 — Memory-efficient foundation-model training
VkD: Improving Knowledge Distillation using Orthogonal Projections
CVPR 2024 — Knowledge distillation
MobileVOS: Real-time Video Object Segmentation
CVPR 2023 — SRUK 2023 Best Paper Award
See my Google Scholar for a full list.
Community Impact
Academic impact:
300+ citations on 1st-author papers
6+ international patents on efficient vision and foundation models.
Peer-reviewed publications at CVPR, NeurIPS, ICCV (Highlight), AAAI, BMVC, and ICASSP.
Open-source impact:
Released code alongside published research.
Contributor to ⚗️ torchdistill and 🤗 HuggingFace PEFT.
Service:
NeurIPS 2024 Outstanding Reviewer
Reviewer for ICLR, NeurIPS, CVPR, ICCV, ECCV, AAAI, PAMI, BMVC, and other ML and computer-vision venues.
Scholarship:
CDT High Performance Embedded and Distributed Systems (HiPEDS) scholarship.
First Class Honours, MEng Electrical and Electronic Engineering, University of Bristol.