Fractal today announced the launch of PiEvolve, an evolutionary agent engine designed for autonomous machine learning and scientific discovery.
Fractal’s PiEvolve ranks among the top-performing agents on OpenAI’s MLE-Bench, a benchmark that assesses how effectively AI systems can solve real-world machine learning challenges. It is the first evaluation agent to exceed 60 percent in overall medal rate and 80 percent in MLE-Bench-Lite performance, thresholds widely recognized as important milestones in autonomous machine learning. These results demonstrate meaningful advances in agent AI capabilities.
PiEvolve also combines strong performance and efficiency. Within a typical 24-hour run, you can achieve results comparable to systems that require longer run times and more compute. Even after 12 hours, it ranks among the top-performing agents and identifies high-quality solutions early.
Unlike traditional machine learning models that are trained once and then deployed, PiEvolve continually tests and improves its unique solution until the available compute budget is fully utilized. Built on a graph-structured search architecture, it integrates inference, code generation, and verification within a unified iterative process. This allows you to tackle complex multivariable optimization problems across supply chains, financial services, and data center operations that are often difficult for static AI systems to perform at scale.
“MLE-Bench is widely recognized as the gold standard for evaluating AI agents on real-world machine learning tasks,” said Srikanth Velamakanni, co-founder, group CEO and vice chairman of Fractal. “PiEvolve’s ranking among the world’s top systems is meaningful validation of our research direction. At Fractal, our goal has always been to empower decision-making for every human in the enterprise. PiEvolve advances that mission by enabling AI systems that continuously improve and deliver measurable business outcomes.”
The main features of PiEvolve are:
- Continuous optimization: Iteratively evolve candidate solutions to improve performance until computational limits are reached.
- Intelligent memory: Use priority-based sampling with decay to avoid local optimization and ensure diverse exploration of solution paths.
- Dual strategy: actively debug weak solutions while improving high-performing solutions to improve overall system performance.
- Production-ready: Includes pause and resume functionality for long-running workloads and seamlessly integrates into enterprise ML pipelines.
- Graph Structure Search: Systematically explore inference, code, and validation loops to generate and refine solutions.
“PiEvolve is ranked as one of the top agents evaluated on OpenAI’s MLE-Bench, which includes systems developed by the world’s leading research institutions. Achieving top-level performance on OpenAI’s MLE-Bench is a significant milestone for our research team and strengthens Fractal’s commitment to advancing enterprise-grade AI,” said Suraj Amonkar, Chief AI Research and Platform Officer at Fractal. says Mr. “We build autonomous systems capable of continuous inference, self-improvement, and evolutionary learning, bringing next-generation agent intelligence to real-world machine learning challenges.”
With the launch of PiEvolve, Fractal strengthens its position at the forefront of enterprise AI innovation. PiEvolve combines continuous improvement, efficiency, and production reliability into one system to help organizations solve complex challenges at scale with confidence. For more information, please visit https://fractal.ai/ai-research/pi-evolve.
Fractal is a global, publicly traded AI company with a vision to power decision-making for every person in the enterprise.
