Fractal Launches PiEvolve, an Evolutionary Agentic Engine for Autonomous Machine Learning and Scientific Discovery

PiEvolve by Fractal ranks among the top-performing agents on OpenAI's MLE-Bench, a benchmark that evaluates how effectively AI systems can solve real-world machine learning challenges. It is the first evaluated agent to surpass 60 percent in Overall Medal Rate and 80 percent in MLE-Bench-Lite performance, crossing thresholds widely viewed as significant milestones in autonomous machine learning. These results mark a meaningful advance in agentic AI capability.

PiEvolve also combines strong performance with efficiency. Within a standard 24-hour run, it delivers results comparable to systems that require longer runtimes and greater compute. Even after 12 hours, it ranks among the top-performing agents, identifying high-quality solutions early.

Unlike traditional machine learning models that are trained once and then deployed, PiEvolve continuously tests and improves its own solutions until the available compute budget is fully utilized. Built on a graph-structured search architecture, it integrates reasoning, code generation, and validation within a unified iterative process. This enables it to tackle complex, multi-variable optimization problems across supply chains, financial services, and data center operations, where static AI systems often struggle to perform at scale.

"MLE-Bench is widely regarded as the gold standard for evaluating AI agents on real-world machine learning tasks," said Srikanth Velamakanni, Co-founder, Group Chief Executive and Vice Chairman, Fractal. "PiEvolve's ranking among the top systems globally is a meaningful validation of our research direction. At Fractal, our ambition has always been to power every human decision in the enterprise. PiEvolve advances that mission by enabling AI systems that continuously improve and deliver measurable business outcomes."

Key features of PiEvolve include:

• Continuous Optimization: Iteratively evolves candidate solutions, improving performance until computational limits are reached.
• Intelligent Memory: Uses priority-based sampling with decay to avoid local optima and ensure diverse exploration of solution paths.
• Dual Strategy: Improves high-performing solutions while actively debugging weaker ones to elevate overall system performance.
• Production-Ready: Includes Pause and Resume capabilities for long-running workloads and integrates seamlessly into enterprise ML pipelines.
• Graph-Structured Search: Systematically explores reasoning, code, and validation loops to generate and refine solutions.

"PiEvolve ranks among the top agents evaluated on OpenAI's MLE-Bench, including systems developed by leading global research labs. Achieving top-tier performance on OpenAI's MLE-Bench is a significant milestone for our research team and reinforces Fractal's commitment to advancing enterprise-grade AI," said Suraj Amonkar, Chief AI Research and Platforms Officer at Fractal. "We are building autonomous systems capable of sustained reasoning, self-improvement, and evolutionary learning, bringing next-generation agentic intelligence to real-world machine learning challenges."

With the launch of PiEvolve, Fractal reinforces its position at the forefront of enterprise AI innovation. PiEvolve brings continuous improvement, efficiency, and production reliability together in a single system, helping organizations solve complex challenges with confidence and scale. For more information, visit https://fractal.ai/ai-research/pi-evolve

About Fractal

Fractal (NSE: FRACTAL) is a publicly listed global enterprise AI company with a vision to power every human decision in the enterprise.

Fractal's suite of businesses includes Asper.ai (enabling interconnected decisions for revenue growth) and Analytics Vidhya (among the world's largest data science communities). Fractal spun out Qure.ai, a global healthcare AI leader enhancing the rapid identification and management of tuberculosis, lung cancer, and stroke. Fractal's dedicated AI Research team is focused on foundational AI advancements, including knowledge-based foundational models, reasoning-based systems, and agentic systems. The team has launched successful products, including MarshallGoldsmith.ai, Vaidya.ai, Kalaido.ai, and the open-source reasoning model Fathom-R1-14B and tool-based reasoning model Fathom-DeepResearch.

Fractal employs over 5,000 professionals across global locations, including the United States, Canada, the UK, the Netherlands, Ukraine, India, Singapore, South Africa, the UAE, and Australia. It has consistently earned recognition as one of India's Best Companies to Work For (Top 100, 2025), a 'Great Workplace' for eight consecutive years, and as one of 'India's Best Workplaces for Women' for five years running by the Great Place to Work® Institute. Fractal was also named a Leader in the 2025 Forrester Wave™ for Customer Analytics Service Providers and earned leadership positions in the Everest Group Peak Matrix Assessment 2025 for AI and Analytics Services, and Information Services Group's 2024 assessments for Data Engineering and Data Science Services.

Photo: https://mma.prnewswire.com/media/2918192/Fractal_Analytics.jpg
Photo: https://mma.prnewswire.com/media/2918193/Fractal_Analytics_Limited.jpg