Haiyue Song: Researcher Profile and Contributions

Pioneering the Future of Language: The Impact and Work of Haiyue Song

In the rapidly evolving landscape of artificial intelligence and natural language processing (NLP), researchers are constantly pushing the boundaries of how machines understand, translate, and generate human communication. Among the rising generation of computer scientists shaping this field is Haiyue Song, an artificial intelligence researcher and engineer whose work spans large language models (LLMs), machine translation, and cross-cultural AI systems.

Academic Foundations and Background

Haiyue Song’s journey into computer science and informatics bridges multiple prestigious institutions. He completed his undergraduate studies at Shanghai Jiao Tong University, earning a degree in Computer Science and Technology alongside a minor in Japanese.

He then moved to Japan to pursue advanced degrees at Kyoto University, where he earned both his Master’s and his Ph.D. in Informatics under the guidance of Professor Sadao Kurohashi. During his doctoral studies, he also served as a research fellow and gained hands-on industry and institutional experience, including work at the National Institute of Information and Communications Technology (NICT).

Core Research Contributions

Song’s research focuses heavily on overcoming linguistic barriers and enhancing the capabilities of AI models, particularly in complex or low-resource scenarios. Some of his most notable contributions include:

• Large Language Models (LLMs) & Continual Pre-Training: His recent work delves into agent harnesses, long-context understanding, reinforcement learning, and optimizing how models undergo continual pre-training (such as his research on distribution vector merging techniques like OptiMer).
• Advanced Machine Translation: Before focusing deeply on generative LLMs, Song made significant waves in neural machine translation (NMT). He explored sub-word segmentation methods (such as SelfSeg), multilingual and multimodal translation scenarios, and structured document translation via reinforcement learning.
• Cultural Awareness and Diversity in AI: Recognizing that AI models often reflect cultural biases, Song has been involved in developing benchmarks—such as the Culturally-diverse Multilingual Visual Question Answering Benchmark (CVQA)—designed to measure and increase cultural awareness and linguistic diversity in multimodal models.

Current Endeavors

Continuing his career in industry and advanced research, Song works as an LLM engineer at Preferred Networks in Tokyo, Japan. There, he focuses on post-training and developing large language models from scratch, helping to bridge theoretical computer science breakthroughs with practical, real-world applications.