Japan-based KDDI and the National Institute of Information and Communications Technology (NICT) started joint research on large-scale language models (LLMs) with the aim of tackling factually incorrect data known as hallucinations.

The research will make use of data from more than 60 billion webpages collected by NICT and technology developed by KDDI Research to control hallucinations, the operator explained in a statement. The company added it will research and develop hallucination suppression technology to match the trends of Japanese general-purpose LLMs.

The pair also will work to enable the handling of multimodal data such as map images and associated building information.

The joint research is the first phase of a project undertaken by the Ministry of Internal Affairs and Communications and NICT to strengthen Japan’s development capabilities for LLMs.

In March, KDDI entered a capital and business alliance with AI start-up Elyza, with the aim of launching a generative AI platform.