Preprint 2026 · Open Full-Stack Release · OpenRSI First Release
Open weights · open gym · open search — the full OpenMLE stack, released
Live · illustrative simulation, not paper trace. OpenMLE-Evo growing a population of executable programs — node size ∝ simulated validation score. Move your cursor to perturb it.
Project Film · 64 seconds
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.
01 · TL;DR
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.
Recursive self-improvement (RSI) requires AI systems that improve the process of building AI (i.e., AI4AI); machine learning engineering (MLE) offers a concrete, executable testbed for studying this capability. We introduce OpenMLE, an open full-stack system for RSI research in MLE, spanning verifiable task environments with execution feedback (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo). On this stack we post-train Frontis-MA1 (35B) as a meta-evolution agent for MLE, aligning post-training and inference around four atomic program-evolution operators (Draft, Improve, Debug, Crossover): the same operators are trained via execution-grounded SFT and RL on data deduplicated against all evaluation benchmarks, then composed into long-horizon search, coupling learning and evolution in a single loop.
On MLE-Bench Lite under a 12-hour per-task budget on one RTX 4090 capped at 12 GB VRAM, Frontis-MA1 (35B) improves Medal Average from 39.39% to 60.61% over its base model with OpenMLE-Evo, and reaches 71.21% with OpenMLE-Evo-Max (benchmark-independent experience priors and asynchronous search), exceeding GPT-5.5 + Codex and approaching GPT-5.6 Sol and the 2.8T Kimi K3. On held-out NatureBench Lite, both components transfer: with the framework fixed, swapping in the trained model raises Match-SOTA from 50% to 70%; with the model fixed, swapping in OpenMLE-Evo raises it from 20% to 50%. We release the model weights and the full OpenMLE stack to enable reproducible research on executable AI4AI toward RSI.
02 · Mechanism Ladder
Verified trajectories return to train the model that generates future evolution.
Artifacts evolve. The improver stays frozen.
Experience flows back into the search.
The improver itself is trained — SFT + RL on verified trajectories.
Each system improves the process that builds its successor.
03 · The Open Full-Stack
Sandbox feedback = training signal = search signal.
MODULE 01 — ENVIRONMENT
MODULE 02 — TRAINING
MODULE 03 — INFERENCE
Trained by OpenMLE. Driving OpenMLE. Evaluated only on third-party benchmarks.
PRODUCT × ENGINE04 · Trainable Atomic Operators
Invoked thousands of times per task — repair, refine, recombine.
① Evolutionary Inference — operator-driven expansion of the solution tree
Scaffold-agnostic → Greedy · abMCTS · AIRA-EVO · OpenEvolve
05 · MLE-Bench Lite · 12 h / task · 1× RTX 4090 · 12 GB VRAM cap · avg of 3 runs
Gains compose everywhere: 30B 34.85 → 53.03 → 66.67 (+18.18pp) · GLM-5.2 59.09 → 66.67 · MiniMax M3 54.55 → 65.15 · Kimi K2.6 59.09 → 66.67 · medal mix shifts toward Gold — matching Kimi K3's Gold rate.
AIRA-Evo baseline · tokens / 12h run
129.3M
Free-form history — every expansion gets more expensive.
OpenMLE-Evo · same checkpoint, seed & budget
75.3M −41.7%
Prompt −50.3% · nodes −12.4% · more new-best updates.
0.9455held-out Human Rank
Debug → Crossover → ConvNeXt leap. 85.0% of the gain from late ops. Best baseline: 0.6303, no medal.
0.8889held-out Human Rank
Memory keeps winning branches, flags a weak ResNet50. 91.9% of the gain from Improve + Crossover. Best baseline: 0.2963, no medal.
06 · RL Training Curves · Frontis-MA1-35B
The rollout reward is noisy by design, while held-out validation and medal count expose the underlying learning trajectory. Hover the chart for exact checkpoint values.
221
RL policy updates logged
0.180 → 0.400
Rollout reward · first to final step
0.404
Peak validation base reward
31 /176
Peak validation medal count · out of 176 tasks
Reward / medals
stitched training step
07 · Meta-Ability & Transfer
NatureBench Lite, controlled: hold the model or the framework fixed — both transfer.
Model transfer · framework fixed · Match-SOTA
Same framework, Frontis-MA1 swapped in.
Framework transfer · model fixed · Match-SOTA
Same model, OpenMLE-Evo search swapped in.
5 / 5
Every modality group improves. No Medal Rate drops.
+2 +4 +1 +4 +3
14 new medals, spread across all five modalities.
24.2 → 34.8
Test medal rate with upper-tail reward shaping (×4.0).
08 · Open-Source Release
R-01
30B & 35B post-trained checkpoints on HuggingFace.
R-02
Audited task inventory + category-specific release artifacts.
R-03
Scheduler + CPU/GPU Docker workers; control / execution / storage separated.
R-04
Execution-grounded SFT data construction + asynchronous RL configs.
R-05
Experience cards, 3-factor selection, operator-conditioned context.
R-06
Full MLE-Bench Lite / NatureBench Lite evaluation & reproduction scripts.
@misc{yang2026frontisma1trainingai4aimodel,
title={Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering},
author={Junlin Yang and Che Jiang and Yu Fu and Tianwei Luo and Can Ren and Weizhi Wang and Kaikai Zhao and Hongyi Liu and Yuxin Zuo and Yuru Wang and Yuchen Fan and Kai Tian and Zhenzhao Yuan and Xiaojian Lin and Li Sheng and Rushi Qiang and Guoli Jia and Xingtai Lv and Ermo Hua and Dianqiao Lei and Youbang Sun and Ning Ding and Bowen Zhou and Kaiyan Zhang},
year={2026},
eprint={2607.28568},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2607.28568},
}