ShinkaEvolve: Evolving new algorithms with LLMs with higher sample-efficiency (sakana.ai)

Sakana AI introduced ShinkaEvolve, an open-source (Apache 2.0) evolutionary code‑optimization framework that uses large language models to evolve new algorithms with dramatically higher sample efficiency. In benchmarks it produced a new state‑of‑the‑art 26‑circle packing solution using only ~150 samples (vs thousands typical of prior work), evolved an effective three‑stage agent scaffold for AIME math problems in 75 generations, improved AtCoder heuristic contest solutions (avg +2.3% score), and discovered a novel Mixture‑of‑Experts load‑balancing loss in 30 generations that reduced inefficient routing and improved downstream accuracy (+1.73% on average). ShinkaEvolve’s efficiency comes from three core innovations: an archive-driven evolutionary loop with parent sampling that balances exploration/exploitation; novelty‑based program rejection sampling (embedding similarity + an LLM “novelty judge”) to avoid evaluating trivial variants; and a bandit-style, task‑dependent LLM prioritization that dynamically picks the most useful model during search. Ablation studies attribute the sample gains to these components working together. By open‑sourcing the code and a WebUI, the team positions ShinkaEvolve as a practical co‑pilot for researchers and engineers to speed algorithm discovery, reduce LLM query costs, and extend evolutionary search to domains—like training methods and agent design—where compute or labeling budgets previously made discovery prohibitive.