Bridging Boundaries: The Life and Work of Dr. Haiyue Song

In an era defined by the rapid evolution of artificial intelligence, breakthroughs rarely happen in a vacuum. They are driven by researchers who sit at intersections—between nations, academic disciplines, and complex technical paradigms. Dr. Haiyue Song, an innovative AI engineer and researcher, embodies this cross-disciplinary spirit. From pushing the frontiers of large language models (LLMs) to advancing neural machine translation (NMT), his journey reflects the borderless nature of modern computer science.

Roots and Academic Foundations

Dr. Song’s academic path established a rigorous foundation in both linguistics and core computer science. He began his undergraduate studies at Shanghai Jiao Tong University, where he earned a Bachelor’s degree in Computer Science and Technology alongside a minor in Japanese from the School of Foreign Languages.

This dual focus on computational logic and human language proved prophetic. It laid the groundwork for his transition to Kyoto University in Japan, where he pursued graduate studies under the mentorship of Prof. Sadao Kurohashi. Immersed in Kyoto University’s renowned intelligence science and technology ecosystem, Dr. Song focused heavily on natural language processing (NLP). He completed his Master’s degree in 2020 and successfully defended his Ph.D. in Informatics in March 2024.

During his doctoral journey, he also served as a prestigious Japan Society for the Promotion of Science (JSPS) Research Fellow (DC1), cementing his reputation as a rising scholar in the international NLP community.

Pioneering Research: From Translation to LLMs

Dr. Song’s research portfolio spans several critical pillars of modern artificial intelligence, bridging traditional text translation mechanics with cutting-edge generative AI:

• Neural Machine Translation & Subword Segmentation: Early in his career, much of Dr. Song’s work tackled the nuances of subword tokenization, cross-granularity alignment, and error span detection. His research—such as DiverSeg and BERTSeg—helped machines better parse and translate complex or low-resource languages by optimizing how words are broken down and understood.
• Large Language Models (LLMs) and Post-Training: As the AI landscape shifted toward generative models, Dr. Song pivoted into the core mechanics of LLMs. His work encompasses agent harnesses, reinforcement learning, long-context handling, and continual pre-training.
• Continual Pre-Training & Model Merging: Addressing how models absorb new information efficiently, his research explores optimal distribution vector merging over traditional data mixing (such as his work on OptiMer) to enhance continual pre-training.

Professional Impact

Dr. Song’s expertise has been sought out by leading research institutions and industry pioneers. Following his time as a technical researcher at the National Institute of Information and Communications Technology (NICT) in Japan, where he contributed heavily to advanced speech translation and multilingual architectures, he transitioned into industry.

Today, he serves as an LLM engineer at Preferred Networks, helping to post-train foundational Japanese language models like PLaMo. Beyond his daily engineering and research duties, he remains deeply active in the global academic community, contributing as an area chair, workshop organizer, and reviewer for premier conferences like EMNLP.

Looking Ahead

Dr. Haiyue Song’s trajectory illustrates the power of a multidisciplinary background. By marrying a deep structural understanding of human linguistics with advanced machine learning architectures, he continues to help shape how machines process, understand, and generate human language across diverse cultures and writing systems.