I am Kazuki Fujii (藤井 一喜), a PhD student in the Rio Yokota Lab at the Institute of Science Tokyo (formerly Tokyo Tech). I am a core contributor to the Swallow Project, leading the development of open Japanese-English bilingual LLMs.
My research lies at the intersection of HPC and Machine Learning, focusing on large-scale distributed training and low-precision optimization (FP8/NVFP4) using Megatron-LM and TransformerEngine. I also study data-centric approaches to improving LLM reasoning, including rewriting pre-training data for Math and Code. My research interests include continual pre-training, Agentic RL, and hardware-aware model architectures.
- [Apr–Sep 2027] I will return to NVIDIA as a Research Intern.
- [Nov 2026] Upcoming: SC 2026.
- [Jun–Sep 2026] I was a PhD intern at NVIDIA Santa Clara, working on SWE-RL and Async RL.
- [Jan 2026] My paper "Rewriting Pre-Training Data Boosts LLM Performance in Math and Code" was accepted to ICLR 2026! 🎉
- 🌐 Website: https://okoge-kaz.github.io/
- 📄 CV: Curriculum Vitae
- 🎓 Google Scholar: Citations Profile
- 💼 LinkedIn: kazuki-fujii