Both technologies—for creation and for applying—are essential for LLM development

My interest in NEC as a potential employer sparked when I saw the news about their development of an original large language model (LLM) “cotomi”. It was when the innovation of LLMs was becoming a hot topic that I started to think about my career path. I studied natural language processing, Vision & Language, and other aspects in the area of AI in university, so I had a strong interest in LLMs. The problem is, LLM development requires massive resources. While I was thinking that research in LLMs would be difficult in academia considering its scale, I came across this news. As I looked further into the company, I learned that NEC has one of the most abundant computational resources in Japan, in addition to its employees who have rich knowledge and experience, and decided to join the company.

Since joining NEC, I have been working on the development of core technologies for the graphic context understanding function (Note 1), which automatically and accurately reads visual documents. The graphic context understanding function automatically converts any data, be it from a Web page, manual, academic paper, slide, poster, or report, into text data that can be taught to an LLM. This was created by adopting NEC’s proprietary learning data and is already commercialized. I believe this can be an essential platform technology for promoting the application of LLMs.

LLM development is currently in a stage where an approach from “how to apply LLMs” is critical. Overseas big techs are using incommensurable resources to develop universal, large-scale LLMs. However, any upgrade to make it smarter won’t immediately make our life better. We also need to develop technologies for effectively applying LLMs in parallel to such upgrades. We are currently setting our eyes on the development of such technologies.

Upgrades in LLMs can make our technologies smarter by simply replacing the existing models. I believe instead of repeating the endless race to upgrade LLMs, it is more useful in the long term to conduct research on such technologies.

Furthermore, NEC has many clients in different industries and sectors, with whom the company has a very close relationship. It therefore has an environment where researchers can hear customers’ opinions and feedback first-hand. This is a huge advantage in tackling the problem of “how to utilize” LLMs.

Producing greater value with systematic research within the organization

What I found interesting about NEC when I joined the company was that research themes were systematized within the organization. In university, our primary goal basically tends to be to do “good research” and contribute to the academic community. At NEC, in addition to doing “good research,” we also need to focus on how to create synergy as an enterprise. There are big goals—to be achieved by a group or as a whole company—and there are technologies that are currently missing or insufficient, but need to be developed or improved in order to achieve those big goals. A researcher then decides to pursue research that improves these technologies, in line with their interests and areas of specialty—your research functions as a component of the bigger goals. For example, the graphic context understanding function, which I am working on, is a technology that is aimed at making a world where LLMs automatically process all kinds of data a reality. The charm of this process is that you create greater value and solutions with your own hands through synergy with other groups. You see the vision of what your research is trying to achieve. The bigger goals can be your guide for deciding the direction of your research theme, as well as give you stronger motivation.

My future goal is to become a researcher who can make a solid impact on society, research community, and customers. In recent years, while the number of research papers continues to increase globally, research achievements that are actually impactful are quite limited. Amid such circumstances, while keeping my focus on essential research, I would like to conduct research that people comment ten years later: “We have this technology now because there was Ueda’s technology.”