Frontis-MA1: Training an AI4AI Model towardsRecursive Self-Improvement in Machine Learning Engineering

Project Film · 64 seconds

See the complete OpenMLE stack

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 Open MP4 ↗

01 · TL;DR

AI that improves the process

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 (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.

0_%_

Medal Average — above GPT-5.5 + Codex (68.18%)

+0_pp_

Post-training gain over base, identical harness

0

Quality-gated executable tasks in OpenMLE-Gym

0

Execution-verified SFT samples

0

Trainable atomic operators = the action space

0_B_

Parameters — Pareto frontier vs trillion-scale systems

---

02 · Mechanism Ladder

When the improver is trained,

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

L1

#### Evolution

Artifacts evolve. The improver stays frozen.

L2

#### Self-Evolution

Experience flows back into the search.

L3

#### Meta-EvolutionTHIS WORK

The improver itself is trained — SFT + RL on verified trajectories.

L4

#### Recursive Self-Improvement

Each system improves the process that builds its successor.

Frontis-MA1 the operator model · the improver TRAINING SIDE OpenMLE-Evo experience-guided long-horizon search INFERENCE SIDE verified trajectories → SFT + RL operator weights → drive search META-EVOLVING

---

03 · The Open Full-Stack

Unit of training \= unit of search.

Sandbox feedback = training signal = search signal.

MODULE 01 — ENVIRONMENT

OpenMLE-Gym_a 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-ERL_execution-grounded SFT + RL_

W ×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-Evo_test-time scaling → test-time learning_

U = λs·s̃ + λΔ·Δ̃ + λn·ν s̃① Δ̃① ν① score leader gain leader 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

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

✦

#### Draft

generate from scratch

↗

#### Improve

refine a parent program

⌗

#### Debug

repair failing code

⑂

#### Crossover

recombine two parents

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

Root · Task & Data 0.61 0.47 Error 0.58 Error 0.32 Error 0.67 0.73 ⋯ ⋯ ⋯

Draft Improve Debug Crossover edge style = operator · node value = sandbox validation score

Inference Loop 1Choose Operator→ 2Select Parent→ 3Generate Code→ 4Sandbox Execute→ 5Evaluate & Rank→ 6Update Database ⟲ iterate

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

A 35B open model at the table

Qwen3.6-35B-A3B base · OpenMLE-Evo

39.39

Frontis-MA1-35B post-trained · OpenMLE-Evo

60.61

Claude Opus 4.8 Claude Code

63.64

GPT-5.5 Codex

68.18

Frontis-MA1-35B OpenMLE-Evo-Max · priors + async parallel search

71.21

GPT-5.6 Sol / Kimi K3 Codex / Claude Code · up to 2.8T

72.73

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.

BRONZE — leaf-classification

#### Late Crossover + Improve jumps

0.9455held-out Human Rank

Debug → Crossover → ConvNeXt leap. 85.0% of the gain from late ops. Best baseline: 0.6303, no medal.

SILVER — mlsp-2013-birds

#### Memory-guided recombination

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

Training signal rises through

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

Rollout reward · EMA 9 Validation reward Validation medals

Left axis · base reward Right axis · medal count Source · qwen36\_rl\_stitched.csv · inlined

---

07 · Meta-Ability & Transfer

It learned how to improve —

NatureBench Lite, controlled: hold the model or the framework fixed — both transfer.

Model transfer · framework fixed · Match-SOTA

50%⟶70_%_

Same framework, Frontis-MA1 swapped in.

Framework transfer · model fixed · Match-SOTA

20%⟶50_%_

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

Models, tasks, training, and search.

R-01

#### Frontis-MA1 weights

30B & 35B post-trained checkpoints on HuggingFace.

R-02

#### OpenMLE Tasks

Audited task inventory + category-specific release artifacts.

R-03

#### Sandbox infrastructure

Scheduler + CPU/GPU Docker workers; control / execution / storage separated.

R-04

#### OpenMLE-ERL

Execution-grounded SFT data construction + asynchronous RL configs.

R-05

#### OpenMLE-Evo

Experience cards, 3-factor selection, operator-conditioned context.

R-06

#### Evaluation harness

Full MLE-Bench Lite / NatureBench Lite evaluation & reproduction scripts.

CITE — BIBTEXCOPY

@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}, }