The Road to Self-Evolving LLMs: From Shinka Evolve to Open-Ended Intelligent Discovery

Introduction: From Evolutionary Algorithms to AI Self-Evolution

Evolutionary algorithms borrow mechanisms from natural selection and genetics to optimize candidate-solution populations through selection, crossover, and mutation. While they naturally support open-ended exploration and global search, traditional implementations are sample-inefficient, often requiring thousands of trials. With the rise of large language models (LLMs), researchers began combining LLM generative and reasoning capabilities with evolutionary search. Google DeepMind's AlphaEvolve was a milestone—evolving algorithms for mathematical optimization, matrix multiplication, and data-center scheduling—but it suffered from very low sample efficiency and a closed-source system.

To address these bottlenecks, Japanese AI startup Sakana AI open-sourced Shinka Evolve ("shinka" meaning "evolution" in Japanese), aiming for unprecedented sample efficiency and open-ended program search toward AI self-evolution.

Shinka Evolve: Breaking the Sample-Efficiency Bottleneck

Core Innovations

Shinka Evolve achieves order-of-magnitude sample-efficiency improvements through three key techniques:

1. Intelligent parent sampling: balances exploration and exploitation when choosing parent programs, avoiding both micro-tweaks of known solutions and aimless random search. 2. Novelty-based rejection sampling: computes embedding similarity between candidate and existing programs and uses the LLM as a "novelty judge," rejecting trivial variations to concentrate evaluation resources on genuinely creative candidates. 3. Task-dependent LLM prioritization: supports multi-model ensembles, using bandit algorithms to adaptively prioritize whichever LLM performs best as evolution progresses.

Validated Results

• Mathematical optimization: On the 26-circle packing problem, Shinka Evolve found a new state-of-the-art solution with only 150 samples, versus thousands of evaluations for AlphaEvolve.
• Agent architecture design: In just 75 generations, it evolved a three-stage agent architecture (diverse expert roles, peer review, synthesis) that significantly outperformed strong baselines on the AIME math benchmark and generalized across years and underlying LLMs.
• Competitive programming: On AtCoder tasks (NP-hard optimization), it improved existing top agents by an average of 2.3%, with one improvement large enough to rank second if entered in competition.
• LLM training optimization: In 30 generations, it discovered a new MoE load-balancing loss function surpassing DeepSeek's "Global LBL," improving efficiency and downstream accuracy on 7 benchmarks. The discovered loss function remained effective on a 5x-larger MoE model, showing broad applicability.

Shinka Evolve is released under the Apache 2.0 license, greatly lowering the barrier for community participation in evolutionary AI research.

From Solving Fixed Problems to Inventing New Ones

A deeper significance is the paradigm shift from solving fixed, human-defined problems toward open-ended evolution, where the system co-evolves the problems themselves. True scientific progress requires problem discovery, not just problem solving.

• In mathematical optimization, Shinka Evolve can arrive at entirely new algorithmic ideas (e.g., combining golden-angle spirals with simulated annealing).
• In agent design, it invented a novel three-stage architecture rather than copying human expert designs.

This move from passive optimization to active invention positions AI as a collaborator—a "co-pilot" for scientists and engineers that proposes new ideas and discovers new problems.

Broader Implications: A New Paradigm for Scientific Discovery

Accelerating Scientific Discovery

Self-evolving AI can serve as an engine for science: proposing new molecular structures, algorithms, or mathematical conjectures and verifying them via evolutionary search. It can escape human cognitive ruts—evolutionary algorithms naturally jump local optima, while LLMs bring cross-domain knowledge and creativity—enabling breakthroughs in directions humans may not have explored.

Human–AI Co-Evolution

AI self-evolution is not about replacing humans but enabling symbiotic co-evolution: humans provide value judgments, ethical guidance, and goals; AI provides massive computational exploration, pattern recognition, and novel ideas. Future research teams may pair human scientists with AI agents, achieving a 1+1>2 effect that combines AI speed and scale with human insight and values.

Conclusion

Shinka Evolve's efficiency revolution and open-source release mark a shift from expensive, closed evolutionary program search toward efficient, open-ended self-evolution—turning AI from a tool that solves given problems into an explorer that invents and solves its own. Challenges remain: ensuring AI-discovered knowledge benefits humanity, preventing reward hacking, and building dynamic evaluation systems. But the trajectory points to a future where AI acts as an intelligent partner, jointly expanding the boundaries of knowledge with humans.