FrontisAI/OpenRSI

OpenRSI – Open Initiative for “AI improving AI”

What it is – OpenRSI (by Frontis AI) is a research‑grade open‑source stack that turns the idea of recursive self‑improvement (RSI) into a concrete engineering problem. It focuses on machine‑learning‑engineering (MLE) as a testbed where an AI system can draft, improve, debug, and recombine code, run the code, observe the results, and learn from that experience. The project bundles:

Component	Role
Frontis‑MA1	A 35‑B (and 30‑B) “AI‑for‑AI” foundation model trained to perform the four atomic program‑evolution operators (Draft, Improve, Debug, Crossover).
OpenMLE	The full‑stack runtime that provides (1) OpenMLE‑Gym – a library for building verifiable MLE tasks and executing them, (2) OpenMLE‑RL – pipelines for supervised‑fine‑tuning (SFT) and execution‑grounded reinforcement learning of the operators, and (3) OpenMLE‑Evo – a long‑horizon evolutionary search engine that repeatedly applies the operators to optimise code.
Datasets	OpenMLE‑Tasks (task definitions) and OpenMLE‑SFT‑Traces (supervised trajectories) hosted on Hugging Face.
Benchmarks	Results are reported on MLE‑Bench Lite (a machine‑learning‑engineering benchmark) and NatureBench Lite (a coding‑agent benchmark derived from Nature‑family papers).

---

Why it matters

• Executable AI‑for‑AI – instead of a static benchmark, the system continuously generates, runs, and learns from real code, making the rate of improvement itself an optimization target.
• Meta‑evolution – the first release stops at “meta‑evolution” (training an improver in a bounded domain) rather than claiming full recursive self‑improvement, which keeps the research claim realistic and reproducible.
• Open, reproducible stack – all code, models, and task data are publicly released (CC‑BY‑NC 4.0 for the repository, model weights on Hugging Face). The README provides step‑by‑step guides for reproducing the training pipeline or running a single search on a NatureBench task.

---

Core concepts

Operator	What it does
Draft	Generates a fresh program from a specification.
Improve	Takes an existing program and refines it using execution feedback (e.g., performance metrics).
Debug	Repairs a program that fails to run or produces errors.
Crossover	Merges two parent programs to create a new candidate.

These operators are learned from execution‑grounded data (SFT traces) and then composed by OpenMLE‑Evo into a long‑horizon search that can run for hours on a single GPU or asynchronously across many GPUs.

---

Getting started (quick‑start)

# Clone the repo
git clone https://github.com/FrontisAI/OpenRSI.git && cd OpenRSI

# Choose a component – for a single NatureBench case:
#   1. Clone the NatureBench repo
#   2. Set up a Python venv and install OpenMLE‑Evo requirements
#   3. Run the provided script (replace placeholders with your model endpoint & API key)

git clone https://github.com/FrontisAI/NatureBench.git
cd OpenMLE-Evo
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
conda env create -f environments/naturebench-local.yml
export PRIMARY_KEY='your‑api‑key'
./.venv/bin/python scripts/run_naturebench_local.py \
  --naturebench-repo ../../NatureBench \
  --conda-env naturebench‑local \
  --model-base-url https://model.example/v1 \
  --model-id served‑model‑name \
  --smoke   # remove for full search

The repository contains separate READMEs for each sub‑module (OpenMLE‑Gym, OpenMLE‑ERL/SFT, OpenMLE‑ERL/RL, OpenMLE‑Evo) that detail environment setup, data preparation, and launch commands.

---

Results snapshot

Evaluation	Setup	Medal Avg / Match‑SOTA
Post‑training gain	Base model → Frontis‑MA1‑35B (search fixed)	39.39 % → 60.61 %
Full system	Frontis‑MA1‑35B + OpenMLE‑Evo‑Max	71.21 %
Model transfer (NatureBench)	Swap in Frontis‑MA1, keep framework	50 % → 70 %
Framework transfer	Swap in OpenMLE‑Evo‑Max, keep base model	20 % → 50 %
These numbers are model‑harness results: they measure how much the learned operators and the search system together improve performance, not a raw model‑only score.

---

Who might use this

• AI‑for‑AI researchers exploring recursive self‑improvement, meta‑learning, or automated software engineering.
• ML engineers who want a sandbox to experiment with AI‑generated code that can be executed and evaluated automatically.
• Benchmark developers looking for a reproducible pipeline that ties task definition, execution, and evaluation together.
• Educators who need an open stack to demonstrate end‑to‑end AI‑driven code optimisation.

---

License & acknowledgements

• Repository code & OpenMLE assets – CC BY‑NC 4.0 (non‑commercial, attribution required). Third‑party libraries keep their own licenses (see NOTICE).
• Model weights are hosted on Hugging Face under the same non‑commercial terms.
• The project thanks the communities behind Qwen, SLIME, Ray, SGLang, Megatron‑LM, Transformers, MLE‑Bench, NatureBench, and many others.

---

Citation

If you build on the work, cite the accompanying arXiv paper:

@misc{yang2026frontisma1trainingai4aimodel,
  title={Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering},
  author={Junlin Yang and et al.},
  year={2026},
  eprint={2607.28568},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2607.28568}
}

---

TL;DR

OpenRSI provides an open, end‑to‑end platform for AI‑driven improvement of AI code. It ships a 35 B “meta‑evolution” model (Frontis‑MA1), a full MLE task‑execution stack (OpenMLE), and datasets that let researchers reproduce the reported gains or extend the system to new domains.