OpenRSI · Frontis-MA1 — Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Project Film · 64 seconds

See the complete OpenMLE stack<br>in one minute.

From executable tasks and execution-grounded learning to evolutionary search and reported results: a visual tour of OpenMLE-Gym, OpenMLE-ERL, OpenMLE-Evo, and Frontis-MA1.

Your browser does not support embedded video. Open the MP4 directly.

H.264 · 3840 × 2160 · no audio<br>Open MP4 ↗

01 · TL;DR

AI that improves the process<br>of building AI

One loop

Draft · Improve · Debug · Crossover — trained with execution-grounded SFT + RL, then composed into long-horizon search. Learning and evolution close.

39.39 → 71.21

Base → post-trained → Evo-Max, MLE-Bench Lite @ 12 h / task on one RTX 4090 capped at 12 GB VRAM. Beats GPT-5.5 + Codex (68.18) , approaching GPT-5.6 Sol and the 2.8T Kimi K3.

100% open

Weights · gym · sandbox · training · search · eval. The full stack, released.

Full abstract

Recursive self-improvement (RSI) requires AI systems that improve the process of building AI<br>(i.e., AI4AI); machine learning engineering (MLE) offers a concrete, executable<br>testbed for studying this capability. We introduce OpenMLE, an open full-stack<br>system for RSI research in MLE, spanning verifiable task environments with execution feedback<br>(OpenMLE-Gym), operator learning (OpenMLE-RL), and<br>long-horizon search (OpenMLE-Evo). On this stack we post-train<br>Frontis-MA1 (35B) as a meta-evolution agent for MLE, aligning post-training and<br>inference around four atomic program-evolution operators<br>(Draft, Improve, Debug, Crossover):<br>the same operators are trained via execution-grounded SFT and RL on data deduplicated against all<br>evaluation benchmarks, then composed into long-horizon search, coupling learning and evolution in a<br>single loop.

On MLE-Bench Lite under a 12-hour per-task budget on one RTX 4090 capped at 12 GB VRAM,<br>Frontis-MA1 (35B) improves Medal Average from 39.39% to 60.61% over its base model with OpenMLE-Evo,<br>and reaches 71.21% with OpenMLE-Evo-Max (benchmark-independent experience priors<br>and asynchronous search), exceeding GPT-5.5 + Codex and approaching GPT-5.6 Sol and the 2.8T Kimi K3.<br>On held-out NatureBench Lite, both components transfer: with the framework fixed, swapping in the<br>trained model raises Match-SOTA from 50% to 70%; with the model fixed, swapping in OpenMLE-Evo raises<br>it from 20% to 50%. We release the model weights and the full OpenMLE stack to enable reproducible<br>research on executable AI4AI toward RSI.

0%<br>Medal Average — above GPT-5.5 + Codex (68.18%)

+0pp<br>Post-training gain over base, identical harness

Quality-gated executable tasks in OpenMLE-Gym

Execution-verified SFT samples

Trainable atomic operators = the action space

0B<br>Parameters — Pareto frontier vs trillion-scale systems

02 · Mechanism Ladder

When the improver is trained,<br>evolution becomes meta-evolution

Verified trajectories return to train the model that generates future evolution.

L1<br>Evolution<br>Artifacts evolve. The improver stays frozen.

L2<br>Self-Evolution<br>Experience flows back into the search.

L3<br>Meta-EvolutionTHIS WORK<br>The improver itself is trained — SFT + RL on verified trajectories.

L4<br>Recursive Self-Improvement<br>Each system improves the process that builds its successor.

Frontis-MA1<br>the operator model · the improver<br>TRAINING SIDE

OpenMLE-Evo<br>experience-guided long-horizon search<br>INFERENCE SIDE

verified trajectories<br>→ SFT + RL

operator weights<br>→ drive search

META-EVOLVING

03 · The Open Full-Stack

Unit of training = unit of search.<br>One stack closes the loop.

Sandbox feedback = training signal = search signal.

MODULE 01 — ENVIRONMENT

OpenMLE-Gyma gym, not a dataset

≈11,000 · Meta Kaggle catalog

3,972 · eligible · 36%

2,839 · executable · 26%

2,240 · quality-gated · 20% ✓

5,758 executable tasks · 3 sources · MLE-Bench excluded

6 structured sandbox feedback modes

MODULE 02 — TRAINING

OpenMLE-ERLexecution-grounded SFT + RL

×4.0

rollout group · bound W clips · entropic top-weighting

26,259 execution-verified SFT samples

Async rollouts · no straggler stalls

MODULE 03 — INFERENCE

OpenMLE-Evotest-time scaling → test-time learning

U = λs·s̃ + λΔ·Δ̃ + λn·ν

s̃①<br>Δ̃①<br>ν①<br>score leader<br>gain leader<br>novel family

Improve

P(gain leader): 10.47% → 17.09%

1 experience card / node → task-global board

On-demand, operator-conditioned memory

Frontis-MA1-30B / 35B

Trained by OpenMLE. Driving OpenMLE. Evaluated only on third-party benchmarks.

PRODUCT × ENGINE

04 · Trainable Atomic Operators

Four operators. One unified<br>action space for code evolution.

Invoked thousands of times per task — repair, refine, recombine.

✦Draft<br>generate from scratch

↗Improve<br>refine a parent program

⌗Debug<br>repair failing code

⑂Crossover<br>recombine two parents

① Evolutionary Inference — operator-driven expansion of the solution tree

Root · Task &...