FrontisAI/OpenRSI

What it solves

OpenRSI aims to make "AI improving AI" (Recursive Self-Improvement) an executable engineering problem. It focuses on creating systems where AI can autonomously improve its own capabilities by iterating through research, execution, and learning, specifically targeting the rate of improvement itself as an optimization target.

How it works

The project implements this through the OpenMLE stack, which focuses on Machine Learning Engineering (MLE) as its first executable domain. The system operates via a loop of search, experience, and training:

• OpenMLE-Gym: Provides verifiable task environments and evaluation tools to create task packages.
• OpenMLE-RL: Uses execution-grounded Supervised Fine-Tuning (SFT) and online Reinforcement Learning (RL) to train specific program-evolution operators.
• Frontis-MA1: A post-trained AI4AI model that acts as a meta-evolution agent, applying learned operators to improve code.
• OpenMLE-Evo: A search runtime that composes these operators into long-horizon search pipelines.

The system utilizes four atomic operators: Draft (creating programs), Improve (refining based on feedback), Debug (repairing failures), and Crossover (recombining programs).

Who it’s for

It is designed for AI researchers and engineers interested in AI4AI, recursive self-improvement, and automated machine learning engineering.

Highlights

• Frontis-MA1: A dedicated AI4AI model (available in 30B and 35B versions) trained for meta-evolution in MLE.
• Execution-Grounded Learning: Aligns post-training and inference around a shared action space of program-evolution operators.
• Verifiable Results: Demonstrates significant gains on MLE-Bench Lite and NatureBench Lite through the combination of trained models and search frameworks.
• OpenMLE Sandbox: A self-hosted distributed code-execution and automatic-evaluation backend.