ShinkaEvolve: Evolving New Algorithms with LLMs, Orders of Magnitude More Efficiently
Sakana AI
Sakana AI open-sourced ShinkaEvolve, a tool that uses LLMs to evolve new algorithms with way fewer tries than before. It cracked a 26-circle packing record in just 150 samples, where older methods needed thousands.
Based on reporting by Sakana AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Evolutionary search has always had a dirty secret: it works, but it's wasteful. Systems like Google's AlphaEvolve can find genuinely novel algorithms by mutating code and keeping what survives, but they often chew through thousands of LLM calls to get there. That's fine if you're Google. It's a nonstarter for a university lab or a solo developer. Sakana AI's new ShinkaEvolve is built specifically to attack that inefficiency, and the numbers in its paper are hard to ignore.
The headline result is a 26-circle packing problem, a classic geometry puzzle, where ShinkaEvolve found a new state-of-the-art arrangement using only 150 samples. The winning approach combined a golden-angle spiral starting layout with gradient refinement and simulated annealing to dodge local optima. Prior evolutionary approaches to similar problems needed orders of magnitude more attempts to get anywhere close.
Sakana didn't stop at geometry. They pointed ShinkaEvolve at agent design for AIME math competition problems, and after just 75 generations it evolved a three-stage scaffold: multiple expert personas tackle a problem, critique each other, then synthesize a final answer. That structure beat strong baselines and, notably, generalized to different years of AIME problems and even different underlying models. In competitive programming, ShinkaEvolve took the best AtCoder heuristic solutions from Sakana's own ALE-Agent and improved them further, adding tricks like advanced caching and a
My take — AI-written commentary, not fact-checked reporting
I'll believe the medicine-and-design pitch when I see it applied somewhere that isn't a benchmark leaderboard, but the sample-efficiency angle here is the real story and it's genuinely underrated. Everyone's obsessed with bigger models; Sakana just showed that smarter search loops can matter just as much, and they open-sourced it under Apache 2.0 instead of gatekeeping it like AlphaEvolve. That's the kind of move that actually moves the field forward instead of just the stock price.
Read more about this at: Sakana AI
