Teaching AI to Upgrade Itself: Inside the OpenMLE Self-Improvement Engine
For years, computer scientists have pursued a tantalizing horizon in artificial intelligence: recursive self-improvement, the threshold where an AI system begins enhancing the very processes used to construct its successors. Now, researchers from Frontis.AI, Horizon Research, and Tsinghua University have taken a major step toward making that loop real.
In a technical report detailing their new platform, OpenMLE, and its flagship 35-billion-parameter model, Frontis-MA1-35B, the team demonstrates an open-source AI capable of writing, executing, debugging, and iteratively refining machine learning code. Tested on complex data-science challenges, Frontis-MA1-35B achieved a top medal-winning rate of 71.21% on MLE-Bench Lite, outperforming larger proprietary setups like GPT-5.5 paired with OpenAI’s Codex.
How AI Learns to Evolve Code
Traditional coding agents often attempt to solve problems in a single pass or rely on simplistic trial-and-error. Frontis-MA1 breaks machine learning engineering down into four distinct program-evolution “operators,” which it practices during post-training:
- Draft: Building an initial, valid machine learning pipeline from raw problem specifications.
- Debug: Diagnosing runtime failures—such as fixing an incompatible PyTorch loss function argument—and restoring execution.
- Improve: Upgrading working code, such as swapping out an older ResNet backbone for a ConvNeXt architecture.
- Crossover: Merging the distinct strengths of two separate parent solutions into a superior hybrid.
To visualize Crossover, imagine one AI attempt excels at parsing irregular tabular inputs, while another attempt designs a powerful graph neural network for physical properties. Instead of discarding one strategy for the other, Frontis-MA1 fuses the robust parser from the first with the neural architecture of the second, producing a unified script that outperforms both predecessors.
Balancing Quality, Progress, and Novelty
Finding optimal solutions across multi-hour coding sessions requires smart compute allocation. The system’s search harness, OpenMLE-Evo, tracks detailed “experience cards” for every candidate script. Rather than blindly chasing whatever code currently holds the highest validation score—a trap that leads agents to polish dead-end ideas—OpenMLE-Evo balances three factors when selecting which branch to expand next:
- Quality: The overall validation score of the script.
- Progress: How much a candidate improved relative to its direct lineage.
- Novelty: Whether the solution introduces an underexplored modeling family.
By combining this experience-guided search with an isolated execution sandbox containing over 5,700 verifiable tasks (OpenMLE-Gym), the model avoids getting stuck in repetitive repair loops.
Broad Transfer and Open Horizons
Remarkably, Frontis-MA1 achieved its benchmark-topping scores operating within a restricted 12-hour budget on a single consumer-grade Nvidia RTX 4090 GPU. The system also demonstrated strong transferability, significantly improving performance on NatureBench Lite—a suite of scientific tasks derived from peer-reviewed papers spanning cellular omics, physical dynamics, and molecular design.
By fully releasing the model weights, dataset pipelines, and sandbox infrastructure, the team provides a reproducible blueprint for “AI for AI.” While fully autonomous recursive self-improvement remains a distant goal, OpenMLE offers a concrete, executable testbed where machine learning agents can reliably learn to build better machine learning agents.
Chat about this paper
To chat about this paper, you'll need a free Gemini API key from Google AI Studio.
Your API key will be stored securely in your browser's local storage.