ELYZA RSI Research to pursue AI self-improvement

ELYZA, a KDDI group company, has launched ELYZA RSI Research to work on recursive self-improvement and AI-driven AI development. This follows broader momentum: frontier model development has sped up as AI takes over parts of the improvement process, with releases shifting from quarterly to monthly. ELYZA's representative director notes that LLM development no longer differentiates on simple tasks, and the focus has moved to AI for Science and Long Horizon tasks, where benchmarks like Humanity's Last Exam have seen scores jump from around 30% a year ago to 68.4% as of September 2026. ELYZA itself has applied its learning infrastructure to LLM-jp's LLM-jp-4, confirming significant performance gains, and has built semi-automated evaluation benchmarks and self-improving harnesses. The new organization aims to integrate these efforts into an autonomous research base, with the first commercial outcome being the ELYZA Voice Agent, which handles contact center operations end-to-end. While RSI is not yet achieved, ELYZA is positioning itself for a future where AI drives development across digital and physical domains, including robot foundation models with KDDI. Whether this accelerates progress or outpaces safety measures remains an open question, and ELYZA says it will disclose research results appropriately.