Sakana AI Unveils ShinkaEvolve: Open-Source LLM-Driven Evolutionary Framework
Sakana AI has unveiled ShinkaEvolve, an innovative open-source framework leveraging large language models (LLMs) in an evolutionary loop to evolve programs for scientific and engineering challenges. Released under the Apache-2.0 license, it's available on GitHub with a WebUI and examples.
ShinkaEvolve tackles sample inefficiency in evolutionary algorithmic approaches by significantly reducing the number of evaluations required. It maintains an archive of evaluated programs and proposes edits via diff edits, full rewrites, and LLM-guided crossovers. The system has demonstrated consistent gains in four domains with small budgets: circle packing, AIME math reasoning, competitive programming, and LLM training.
Periodically, ShinkaEvolve produces a meta-scratchpad that summarises recent successful strategies, feeding them back into prompts to accelerate later generations. This has led to improvements in AIME scaffolds under strict query budgets, ALE-Bench solutions, and the discovery of a new MoE load-balancing loss that enhances perplexity and downstream accuracy.
In a notable achievement, ShinkaEvolve has reproduced and surpassed the circle-packing result of AlphaEvolve with orders-of-magnitude fewer samples. It targets wasteful exploration in code-evolution systems with adaptive parent sampling, novelty-based rejection filtering, and bandit-based LLM ensembling.
ShinkaEvolve, released under the Apache-2.0 license with a research report and public code, drastically cuts the number of evaluations needed to reach strong solutions. It has set a new state-of-the-art configuration on the circle-packing benchmark using ~150 program evaluations. The open-source framework is poised to accelerate research and development in scientific and engineering problem-solving.
