What is recursive self-improvement, or RSI?

It is AI developing AI itself, an improvement loop that ELYZA says is now driving frontier progress. ELYZA's CEO describes it as work formerly done by people being shifted to AI.

What results has ELYZA reported from RSI?

For inquiry handling, task completion improved from about 47% to about 61%. ELYZA also cut inference costs by about 67% at the same accuracy.

How fast are new AI models appearing now?

Soneoka says new models appeared every three months through 2025. Now, he says, new models keep coming out every month.

Large Language ModelsAI Business & IndustryITmedia AI＋2m agoOct 6, 2026, 16:00 JST

ELYZA RSI Research targets self-improving AI

Send on LINE

3 Key Points

What happened
KDDI's ELYZA said on October 2 it set up "ELYZA RSI Research" to study recursive self-improvement, and CEO Yuya Soneoka said RSI work raised inquiry-handling task completion from about 47% to about 61% and cut inference costs by about 67%.
Why it matters
The gains suggest RSI is a practical lever, not just theory.
What to watch
The test is whether these gains scale beyond inquiry handling, where completion rose from about 47% to about 61%. Watch for fresh monthly releases as evidence the RSI loop holds at frontier complexity.

WHO IT HITSJapanese enterprises and public-sector buyers weighing self-hosted AI are the clearest beneficiaries, since ELYZA frames its work as an alternative to foreign frontier AI. R&D leaders planning model refreshes may face shorter replacement cycles, and cost planners get a concrete datapoint: about 67% lower inference cost at similar accuracy, per the company.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

ELYZA, based in Bunkyo Ward, Tokyo and part of the KDDI group, says it began RSI research before formally establishing ELYZA RSI Research. That earlier work produced the inquiry-handling numbers cited at the October 2 briefing, where CEO Yuya Soneoka framed RSI as the mechanism behind rapid frontier releases. He points to complex tasks, such as AI for Science in drug discovery, chemistry and math, and long multi-step tasks AI completes autonomously, including one case where about 10,000 AI agents worked on an unsolved math problem. Soneoka also says LLM performance on simple tasks like summarization peaked around 2024, leaving little difference between models. Taken together, the company is positioning RSI as a route to an alternative to foreign frontier AI, and the stakes hinge on whether the reported gains hold outside inquiry handling. If they do, ELYZA's results would suggest Japanese developers can compete on efficiency rather than scale, though the evidence so far comes from one task type.

FAQ

What is recursive self-improvement, or RSI?: It is AI developing AI itself, an improvement loop that ELYZA says is now driving frontier progress. ELYZA's CEO describes it as work formerly done by people being shifted to AI.
What results has ELYZA reported from RSI?: For inquiry handling, task completion improved from about 47% to about 61%. ELYZA also cut inference costs by about 67% at the same accuracy.
How fast are new AI models appearing now?: Soneoka says new models appeared every three months through 2025. Now, he says, new models keep coming out every month.

ITmedia AI＋Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.