As recursive self-improvement (RSI), the process of autonomously improving AI models, draws attention, ELYZA announced the launch of ELYZA RSI Research in Tokyo on October 2, 2026. The company also presented ELYZA Voice Agent, a voice-conversation AI agent using an autonomous research platform, and its work on physical AI.

Founded in September 2018, ELYZA is an AI startup spun out of the University of Tokyo’s Matsuo Laboratory. It supports the practical use of generative AI and the development of large language models for companies and industries, and became part of the KDDI Group in April 2024. CEO Yuya Soneoka said ELYZA has pursued LLM research and development since 2019 and has supported more than 50 companies with generative AI research, development, and use. He also introduced ELYZA Works, a tool for businesses.

Soneoka said LLMs had reached a ceiling in accuracy on simple tasks by 2024. Research as of September 2026 is focused on “AI for Science,” which applies AI to areas such as drug discovery, materials, and mathematics, and “Long Horizon” tasks that require completing multiple steps over several hours or days. The Artificial Analysis Intelligence Index (AAII) benchmarks focus on advanced reasoning, mathematical proofs and logical development, coding, completing practical work, and reliability. New models are appearing almost weekly, and they can solve 60% of difficult problems. There are also cases in which coordinated AI agents have solved open mathematical problems.

OpenAI released GPT-6 Astra on September 3, 2026, with a major improvement in its “Computer Use” ability to operate a PC and complete tasks across multiple steps. Chinese companies are also developing and open-sourcing foundation models. Soneoka said that, in his view, open models from China are appearing about two months behind. Building AI agents requires not only foundation models but also “harnesses,” supporting systems for advanced processing, and evaluation systems to measure quality. At present, people develop those systems and improve their accuracy.

Since early 2026, “AI developing AI” has accelerated model progress and increased the frequency of frontier model releases. RSI envisions shifting the improvement loop, in which the results of upgrades are fed back into the system, from people to AI, and automating the loop so AI can improve itself. At the same time, governments are intervening and regulating, and some frontier labs are slowing development. Soneoka said Japan should pursue full-stack sovereign AI development and RSI research, warning that risk management and regulation could restrict public access to leading models or keep them from being released.

As recursive self-improvement (RSI), the process of autonomously improving AI models, draws attention, ELYZA announced the launch of ELYZA RSI Research in Tokyo on October 2, 2026. The company also presented ELYZA Voice Agent, a voice-conversation AI agent using an autonomous research platform, and its work on physical AI.

Founded in September 2018, ELYZA is an AI startup spun out of the University of Tokyo’s Matsuo Laboratory. It supports the practical use of generative AI and the development of large language models for companies and industries, and became part of the KDDI Group in April 2024. CEO Yuya Soneoka said ELYZA has pursued LLM research and development since 2019 and has supported more than 50 companies with generative AI research, development, and use. He also introduced ELYZA Works, a tool for businesses.

Soneoka said LLMs had reached a ceiling in accuracy on simple tasks by 2024. Research as of September 2026 is focused on “AI for Science,” which applies AI to areas such as drug discovery, materials, and mathematics, and “Long Horizon” tasks that require completing multiple steps over several hours or days. The Artificial Analysis Intelligence Index (AAII) benchmarks focus on advanced reasoning, mathematical proofs and logical development, coding, completing practical work, and reliability. New models are appearing almost weekly, and they can solve 60% of difficult problems. There are also cases in which coordinated AI agents have solved open mathematical problems.

OpenAI released GPT-6 Astra on September 3, 2026, with a major improvement in its “Computer Use” ability to operate a PC and complete tasks across multiple steps. Chinese companies are also developing and open-sourcing foundation models. Soneoka said that, in his view, open models from China are appearing about two months behind. Building AI agents requires not only foundation models but also “harnesses,” supporting systems for advanced processing, and evaluation systems to measure quality. At present, people develop those systems and improve their accuracy.

Since early 2026, “AI developing AI” has accelerated model progress and increased the frequency of frontier model releases. RSI envisions shifting the improvement loop, in which the results of upgrades are fed back into the system, from people to AI, and automating the loop so AI can improve itself. At the same time, governments are intervening and regulating, and some frontier labs are slowing development. Soneoka said Japan should pursue full-stack sovereign AI development and RSI research, warning that risk management and regulation could restrict public access to leading models or keep them from being released.