ShinkaEvolve in Action: How a Human-AI Partnership Conquered a Coding Challenge (sakana.ai)

Team Unagi’s first-place finish at the 2025 ICFP Programming Contest was powered by a human-AI partnership: the team fed their manually written Rust SAT-encoding solver into ShinkaEvolve, Sakana AI’s evolutionary code-optimization framework, which used an ensemble of large language models to iteratively mutate and evaluate programs with solver execution time as the fitness metric. Over ~320 trials costing about $60, ShinkaEvolve found a series of changes that cut execution time by up to 10x (e.g., 127s → 13s on a 24-room instance, and 2.86s → 0.44s on an 18-room instance), enabling the team to solve 30-room cases previously out of reach and materially boost their contest score. Crucially, the wins weren’t just opaque speedups: the AI discovered a higher-level encoding trick—introducing an auxiliary variable that shifts representation from “door-to-door” connections to “vertex-first” connectivity—that made the SAT solver’s search far more efficient. The framework maintains an archive of evaluated programs, generates candidates with LLMs, and optimizes for runtime, producing interpretable insights the team then re-applied manually to other problems. This case demonstrates how AI can perform targeted, high-volume search to improve algorithm encodings and yield generalizable design principles, pointing to a practical human+AI workflow for accelerating complex algorithm engineering.