Fractal's PiEvolve AI agent outperforms competitors in machine learning tasks
Fractal just dropped PiEvolve, its new evolutionary, agentic engine designed for autonomous machine learning and evaluated on OpenAI's MLE-Bench.
Across 75 competitions, it's already making waves—scoring medals in over 60% of challenges and earning an impressive 80%+ on MLE-Bench-Lite.
How PiEvolve works
PiEvolve stands out by using a smart graph-based system that combines reasoning, code generation, and validation in constant feedback loops.
It keeps improving itself with intelligent memory tricks and can handle complex problems where older AIs get stuck.
Unlike models that learn once and stop, PiEvolve is always testing and evolving to get better.
Real-world applications and flexibility
In head-to-head tests, PiEvolve ranked among the top AI agents from around the world—delivering competitive results in both short (12-hour) and longer (24-hour) runs.
Its pause-and-resume feature also makes it super practical for real-world uses like supply chains or finance, giving businesses more flexibility with their machine learning projects.