Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab

-->

-->

The Next Paradigm of Artificial Intelligence

As the world enters the era of artificial intelligence, Japan has a unique opportunity to reclaim its position at the frontier of global innovation. However, to achieve global leadership in AI and scientific discovery, we cannot simply stick to the conventional approach of brute-forcing monolithic models. We must leapfrog the current paradigm.

History shows us how Japan’s historical dominance in manufacturing was not achieved through abundant natural resources but by fundamentally redesigning the institution of the factory floor. Through the philosophy of continuous, compounding self-improvement, Japan created systems that achieved more with less.

This same principle applies to intelligence itself. Human cognition did not emerge from limitless resources; it was forged through the open-ended, compounding process of evolution operating under strict constraints. Similarly, building AI in Japan provides the ultimate design constraint. Rather than relying on brute-force scaling, we are driven to pursue elegance, adaptability, and autonomy.

To achieve this, at Sakana AI, we are building open-ended, adaptive architectures that collectively self-improve. Just as biological evolution innovates endlessly by building upon past discoveries, our AI systems must transition from being static tools to autonomous researchers.

Sakana AI is one of the earliest labs developing Recursive Self-Improvement (RSI) technology using modern foundation models. Today, we are proud to announce the formal establishment of the Sakana AI RSI Lab , a dedicated research group within Sakana AI, tasked with redesigning the AI development process itself with AI.

By transitioning from static, human-led R&D to autonomous, self-improving intelligence engines, we are turning constraints into our greatest compounding advantage. We are building the definitive architecture for the next frontier of AI.

Our Lineage: Pioneering the Foundations of RSI

While the industry increasingly speculates about the future theoretical potential of self-improving AI, Sakana AI has spent the last two years shipping practical milestones towards making this a reality.

The RSI Lab does not start from scratch; it builds upon a rich chronological portfolio of breakthrough research that has systematically shifted the industry from hand-designed heuristics to autonomous, evolutionary optimization loops.

The chronological portfolio below documents our work:

Sakana AI’s RSI Research

LLM-Squared (2024): Developed in collaboration with Oxford and Cambridge, this framework pioneered AI-driven automation to let LLMs invent better ways to train LLMs (LLM²). It yielded DiscoPOP, a state-of-the-art preference optimization algorithm discovered and written entirely by an LLM through a generational evolutionary loop. For us, this work sparked an “AI² paradigm shift”: AI models have become powerful enough to start conducting research to improve themself.

The Darwin Gödel Machine (2025): Developed in collaboration with researchers at the University of British Columbia (UBC), DGM enables open-ended continuous self-improvement by maintaining an evolving lineage of agent variants that autonomously rewrite their own codebase. DGM automatically more than doubled its baseline software-engineering performance on SWE-bench, driving a 30 percentage point absolute improvement.

ShinkaEvolve (2025): An open-source framework demonstrating unprecedented sample-efficiency in program evolution for scientific discovery. Utilizing adaptive sampling and novelty filtering, it solved complex optimization problems using only 150 samples and successfully generated a novel load-balancing loss function that improves Mixture-of-Experts (MoE) models.

ALE-Agent (2025): Our milestone optimization agent that secured 1st place out of 804 human participants in the AtCoder Heuristic Contest 058. Leveraging massive inference-time scaling and a self-learning mechanism that extracts insights from trial-and-error failures, it autonomously derived a novel algorithm that outperformed human experts.

Digital Red Queen (2026): A collaboration with MIT establishing open-ended adversarial coevolution within the Turing-complete sandbox of Core War. Driven by an evolutionary arms race where LLMs authored competing code, the system triggered the autonomous emergence of complex software strategies and demonstrated a remarkable form of convergent evolution. This adversarial sandbox lays the foundation for applying RSI to cybersecurity, modeling how autonomous agents can continuously co-evolve to discover, exploit, and patch vulnerabilities in a dynamic algorithmic arms race.

The AI Scientist (2024–2026): Our landmark system capable of fully automated, open-ended scientific discovery, from generating ideas, running experiments, to writing full papers, and executing peer reviews. This research was recognized...