- [14 Jan 2026] PiEvolve is officially Rank-1 on OpenAI's MLE-Bench benchmark. Check out: openai/mle-bench
- [06 Jan 2026] PiEvolve achieves SOTA performance on OpenAI MLE-Bench and is among top-3 even with 50% less compute time than official runtime of 24 hours.
- [Nov, 2025] Re-invented PiML to develop PiEvolve, pushing the boundaries of autonomous scientific discovery via an evolutionary agentic framework.
- [03 Jun 2025] Published "PiML: Automated Machine Learning Workflow Optimization using LLM Agents" at AutoML'25, New York, USA.
- [April, 2025] Developed PiML - a framework for autonomous optimisation of machine learning workflows.
Pi-Evolve is a persistent evolutionary framework for autonomous machine learning and scientific discovery. Pi-Evolve reframes algorithmic optimization as graph-structured evolutionary search over solution trajectories, where each node encodes reasoning, executable code, and empirical validation, and edges capture parent–child lineage across iterative improvement cycles. Unlike prior evolutionary agents that operate over unstructured populations, Pi-Evolve evolves a directed acyclic graph (DAG) of NodeChains, enabling explicit reasoning over historical dependencies and fine-grained credit assignment across evolving solution paths. At its core, Pi-Evolve employs a probabilistic action selection policy;balancing exploration and exploitation via adaptive priority scoring; to orchestrate the operations of specialized LLM-driven agents.
PiEvolve maintains a growing directed acyclic graph (DAG) of solution states. Each node represents a concrete execution step such as drafting a new approach, refining an existing pipeline, or debugging a failure. Edges capture how one solution evolves from another. A sequence of connected nodes forms a solution chain, representing a coherent line of reasoning and experimentation. Multiple solution chains coexist, allowing PiEvolve to explore alternatives in parallel, preserve failed or partial attempts, and reuse intermediate insights instead of discarding them.
At every iteration, PiEvolve selects a high-level action that determines how the solution graph is expanded. Actions include drafting new solutions, improving or debugging existing ones, merging ideas across chains, and applying small mutations to escape local optima. Each action is executed by an agent that produces reasoning, generates executable code, runs it, and records the outcome. The result is immediately fed back into the graph, closing the loop between reasoning, execution, and learning.
Action selection is governed by a probabilistic policy conditioned on the current solution graph. Each active solution chain maintains a priority score that reflects both its quality and recent progress. When overall performance is weak, PiEvolve favors exploratory actions that introduce new ideas. As strong solution chains emerge, the system increasingly focuses on refinement and debugging. This adaptive balance prevents premature convergence while maintaining steady improvement.
All reasoning traces, code, execution results, and scores are stored persistently in the solution graph. Poorly performing or stagnant chains are automatically deprioritized or terminated, allowing computation to focus on more promising directions. By combining long-term memory with iterative feedback, PiEvolve enables continuous, robust optimization without restarting the search or relying on a single best guess.
PiEvolve achieves state-of-the-art performance across difficulty tiers while maintaining competitive runtime efficiency. As shown in Table X, PiEvolve attains the highest overall success rate (61+%) among all evaluated agents, outperforming strong baselines such as Famou-Agent 2.0 and ML-Master 2.0 across Low/Lite, Medium, and aggregate settings. Notably, PiEvolve matches the best reported performance on High-difficulty tasks while exhibiting substantially stronger gains in Medium regimes, indicating robust optimization beyond easy cases.
From an efficiency perspective, PiEvolve demonstrates a favorable performance–compute tradeoff. Despite operating under a standard 24-hour budget, its results are competitive with or superior to agents requiring longer runtimes (e.g., Neo at 36h). Moreover, when considering shorter-horizon performance, PiEvolve ranks 3rd among all agents evaluated at a 12-hour runtime, highlighting its ability to discover high-quality solutions early in the search process. These results suggest that PiEvolve’s persistent evolutionary control and memory-driven refinement enable both scalable performance and efficient use of compute, outperforming prior agentic and AutoML baselines such as AIDE by a large margin.
If you use Pi-Evolve agent, cite us:
@misc{PiEvolve,
title={Pi-Evolve: Long-Horizon Evolutionary Optimization for Autonomous Scientific Discovery},
author={Sai Kiran Botla and Kirubanath Sankar and Abhishek Chopde and Fardeen Pettiwala},
year={2025}
}The Project Pioneer team at Fractal AI Research Lab investigates the next generation of autonomous intelligence by developing agentic systems capable of sustained reasoning, self-improvement, and evolutionary learning over complex machine learning tasks.
Our Published Work:
- "PiML: Automated Machine Learning Workflow Optimization using LLM Agents" [AutoML'25, NY, USA]
This project is licensed under the MIT License. See the LICENSE file for details.
- GitHub Issues: Submit Issue