ShinkaEvolve
ShinkaEvolve
ShinkaEvolve is an open-source evolutionary program optimization framework from Sakana AI that uses LLMs as mutation operators. Published in 2025, it addresses the key limitations of prior closed-source systems like AlphaEvolve: poor sample efficiency and lack of reproducibility. ShinkaEvolve achieved state-of-the-art results across four domains using orders of magnitude fewer evaluations than alternatives.
Key Innovations
- Power-law parent sampling. Balances exploitation (improving top solutions) with exploration (trying novel approaches). Programs are ranked by fitness and selected with probability proportional to rank raised to a tunable exponent.
- Novelty rejection-sampling. Uses code embeddings to filter out mutations too similar to existing programs in the archive. If cosine similarity exceeds a threshold, an LLM judge further assesses whether the mutation is meaningfully different.
- Bandit-based LLM ensemble. Uses UCB1 to dynamically allocate mutation work across multiple LLMs (GPT, Claude, Gemini, DeepSeek), rewarding models that produce higher-improvement mutations.
Results
- Circle packing — new state-of-the-art solution in 150 samples (vs. thousands for AlphaEvolve)
- AIME math reasoning — evolved a multi-expert agent scaffold that generalizes across LLM backends and unseen problems
- ALE-Bench — improved existing competitive programming solutions by ~2.3%, achieving 2nd place on one AtCoder task
- MoE load balancing — discovered a novel loss function that adds adaptive regularization for under-utilized experts, improving on the global-batch loss used by major open LLMs
Connection to Existing Wiki
ShinkaEvolve extends the Autonomous Code Optimization paradigm: where autoresearch optimizes a single codebase with benchmarks and tests, ShinkaEvolve maintains a population of programs and evolves them using LLM-guided mutation and selection. Both are forms of agent-driven automated optimization. The MoE result is particularly notable — it shows evolutionary search can discover improvements to the training process of LLMs themselves.