Kazuki Fujii

藤井一喜 Kazuki Fujii is a PhD student in the Rio Yokota Lab at the Institute of Science Tokyo (formerly Tokyo Tech). He is a core contributor to the Swallow Project, leading the development of open Japanese-English bilingual LLMs.

His research sits at the intersection of HPC and Machine Learning, with a focus on large-scale distributed training and low-precision optimization (FP8/NVFP4) using Megatron-LM and TransformerEngine. He also researches data-centric approaches to improve reasoning capabilities in LLMs. His recent work on rewriting pre-training data for Math and Code was accepted to ICLR 2026.

From Jun to Sep 2026, he was a PhD intern at NVIDIA Santa Clara, where he conducted research on SWE-RL and Async RL. He will return to NVIDIA as a Research Intern from Apr to Sep 2027.

Contact: kazuki.fujii [at] rio.scrc.iir.isct.ac.jp

CV Publications Blog Blog(Japanese)

Research Interests

2023 2024 2025 2026

Distributed Training

Continual Pre-Training

Low Precision Training

Data Improvement

LLM RL

Agentic RL

Hardware aware model architecture

Education

2026 – 2029 PhD in CS, Institute of Science Tokyo Adv: Rio Yokota · Expected Mar 2029

2024 – 2026 MS in CS, Institute of Science Tokyo Adv: Jun Sakuma

2020 – 2024 BS, Tokyo Institute of Technology Adv: Rio Yokota

2016 – 2019 Azabu High School (麻布高等学校)