Haiyue Song: Biography, Research, and Career Overview

The Man Behind the Code: Exploring the Work and Journey of Haiyue Song

In the rapidly evolving landscape of artificial intelligence, breakthroughs are rarely born in a vacuum. Behind every advanced large language model (LLM), multilingual translation engine, and sophisticated neural network framework lies the meticulous work of dedicated researchers. One prominent name making waves in modern AI research is Haiyue Song, an engineer and researcher whose work bridges the gap between massive computational models and cross-lingual communication.

Roots and Nomenclature

Born with a name that carries a sense of profound movement, Haiyue Song (宋海越, Sòng Hǎiyuè) has a personal background deeply tied to history and geography. His family name, Song, traces its lineage back to the ancient State of Song, while his given name, Haiyue, translates literally to “crossing the sea” (Hai meaning sea, and Yue meaning to cross). Fittingly, his academic and professional life has been defined by literal and metaphorical border-crossing—navigating multiple languages, cultures, and technological domains.

Academic Excellence and Foundations

Song’s path into cutting-edge computer science began with rigorous multidisciplinary training. He completed his undergraduate studies at Shanghai Jiao Tong University, where he pursued a dual focus in Computer Science and Technology alongside a minor in Japanese.

This early immersion in linguistics and computer science laid the foundation for his graduate studies at Kyoto University in Japan. Under the guidance of prominent researchers like Professor Sadao Kurohashi, Song dove deep into informatics and artificial intelligence, earning his Master’s degree in Intelligence Science and Technology followed by a Ph.D. in Informatics in 2024. During this time, he also served as a prestigious Japan Society for the Promotion of Science (JSPS) Research Fellow, cementing his status as an emerging talent in computational linguistics.

Core Contributions to AI and Machine Translation

While many modern AI researchers focus exclusively on high-resource languages like English, a major hallmark of Song’s research is his dedication to low-resource, multilingual, and multimodal machine translation, as well as the enhancement of large language models. Some of his notable contributions include:

• Subword Regularization and Segmentation: Tackling how models break down words into smaller chunks, Song has researched advanced tokenization methods (such as SelfSeg and SubMerge) to improve translation accuracy across diverse language structures.
• Continual Pre-Training and Alignment: Working on techniques like optimal distribution vector merging (OptiMer), his research explores how to update and adapt pre-trained language models efficiently without losing core capabilities.
• Document and Speech Translation: Moving beyond simple text sentences, his work expands into structured document translation and speech translation frameworks (collaborating with institutions like Japan’s National Institute of Information and Communications Technology, or NICT).

Looking Forward

Today, Haiyue Song applies his expertise as an LLM engineer at Preferred Networks in Tokyo, focusing on post-training, agent harnesses, long-context capabilities, and reinforcement learning for next-generation language models.