SB Intuitions Corp. (“SB Intuitions”) and the National Institute of Information and Communications Technology (“NICT”) announced today the launch of joint research on safety technology for high-performance large language models (LLMs), which began on February 18, 2026.
In this joint research, NICT’s high-quality linguistic resources and information processing technology accumulated over many years, along with its expertise in safety technologies such as the detection and suppression of inappropriate outputs, will be combined with SB Intuitions’ advanced model-building expertise gained through the development of “Sarashina,” a Japanese language homegrown LLM. By maximizing the use of the latest AI computing infrastructure provided by SoftBank Corp., the two organizations aim to develop high-performance LLMs equipped with even higher levels of safety.
Background
While the convenience of generative AI has attracted significant attention with its rapid spread, ensuring safety has become an imperative concern. This includes addressing risks such as “hallucinations,” where AI generates information inconsistent with facts, as well as inappropriate expressions and potential copyright infringements. In particular, there is a growing demand from domestic corporations and public institutions to use “trustworthy AI” that reflects Japanese legal systems and ethical values, in addition to possessing superior Japanese language capabilities.
Against this background, NICT, which possesses some of Japan’s most extensive linguistic resources, and SB Intuitions, which promotes advanced LLM development using large-scale computing infrastructure, have formed an integrated R&D system. This collaboration will advance the entire process from data construction and model training to evaluation and safety assurance.
Research Overview
Development of alignment and evaluation metrics for improving LLM safety:
Development of technologies known as alignment to adjust LLMs and ensure that they operate appropriately in accordance with human values and ethics, and development of metrics to measure their safety.
Development of guardrail technology to detect inappropriate expressions:
Research on guardrail technology to detect inappropriate expressions in both input sentences and the outputs generated by the LLM, enabling control through filtering or blocking.
Roles of Each Organization
SB Intuitions
- Implementation of base model development and training for high-performance LLMs.
- R&D of safety assurance technologies, including alignment and guardrail technology.
NICT
- Provision of linguistic resources and expertise accumulated in information processing technology over the years, along with strengthening of data to improve LLM safety.
- Development of fundamental technologies for LLM safety assurance and evaluation, such as the detection and suppression of inappropriate outputs.
Company Profile
SB Intuitions Corp.
As a subsidiary of SoftBank Corp., SB Intuitions focuses on research and development of generative AI and related services centered on the “Sarashina” series of Japanese-specialized large language models, aiming to achieve the best performance in Japanese among domestic LLMs. Leveraging one of Japan's largest computing infrastructures and stringent domestic data management, the company has released high-performance models such as “Sarashina2-70B” and the multimodal model “Sarashina 2-Vision,” accelerating the social implementation of generative AI through industry-academia collaboration.
National Institute of Information and Communications Technology
As Japan's sole National Research and Development Agency specializing in the field of information and communications technology, NICT is charged with promoting ICT sector as well as research and development in ICT, which drives economic growth and creates an affluent, safe and secure society.
sbintuitions.co.jp