{"slug": "frontis-ma1-open-35b-model-toward-recursive-self-improvement", "title": "Frontis-MA1: Open 35B model toward recursive self-improvement", "summary": "Frontis AI released Frontis-MA1, a 35B open-weights model, along with the full OpenMLE stack, claiming it improves Medal Average on MLE-Bench Lite from 39.39% to 60.61% with OpenMLE-Evo and to 71.21% with OpenMLE-Evo-Max, surpassing GPT-5.5 + Codex (68.18) and approaching GPT-5.6 Sol and Kimi K3. The system integrates execution-grounded training and evolutionary search for recursive self-improvement in machine learning engineering.", "body_md": "Preprint 2026 · Open Full-Stack Release · OpenRSI First Release\n\nOpen weights · open gym · open search — the full OpenMLE stack, released\n\nLive · illustrative simulation, not paper trace.\n**OpenMLE-Evo** growing a population of executable programs —\nnode size ∝ simulated validation score. Move your cursor to perturb it.\n\nProject Film · 64 seconds\n\nFrom 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.\n\n01 · TL;DR\n\nOne loop\n\n**Draft · Improve · Debug · Crossover** — trained with execution-grounded SFT + RL, then composed into long-horizon search. Learning and evolution close.\n\n39.39 → 71.21\n\nBase → 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.\n\n100% open\n\nWeights · gym · sandbox · training · search · eval. **The full stack, released.**\n\nRecursive 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.\n\nOn 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.\n\n02 · Mechanism Ladder\n\n**Verified trajectories return to train the model that generates future evolution.**\n\nArtifacts evolve. The improver stays frozen.\n\nExperience flows back into the search.\n\nThe improver itself is trained — SFT + RL on verified trajectories.\n\nEach system improves the process that builds its successor.\n\n03 · The Open Full-Stack\n\n**Sandbox feedback = training signal = search signal.**\n\nMODULE 01 — ENVIRONMENT\n\nMODULE 02 — TRAINING\n\nMODULE 03 — INFERENCE\n\nTrained by OpenMLE. Driving OpenMLE. Evaluated only on third-party benchmarks.\n\nPRODUCT × ENGINE04 · Trainable Atomic Operators\n\nInvoked **thousands of times per task** — repair, refine, recombine.\n\n① Evolutionary Inference — operator-driven expansion of the solution tree\n\nScaffold-agnostic → **Greedy** · **abMCTS** · **AIRA-EVO** · **OpenEvolve**\n\n05 · MLE-Bench Lite · 12 h / task · 1× RTX 4090 · 12 GB VRAM cap · avg of 3 runs\n\nGains compose everywhere: 30B **34.85 → 53.03 → 66.67** (+18.18pp) · GLM-5.2 **59.09 → 66.67** ·\nMiniMax M3 **54.55 → 65.15** · Kimi K2.6 **59.09 → 66.67** ·\nmedal mix shifts toward **Gold** — matching Kimi K3's Gold rate.\n\nAIRA-Evo baseline · tokens / 12h run\n\n129.3M\n\nFree-form history — every expansion gets more expensive.\n\nOpenMLE-Evo · same checkpoint, seed & budget\n\n75.3M −41.7%\n\nPrompt **−50.3%** · nodes **−12.4%** · **more** new-best updates.\n\n0.9455held-out Human Rank\n\nDebug → Crossover → ConvNeXt leap. **85.0%** of the gain from late ops. Best baseline: 0.6303, no medal.\n\n0.8889held-out Human Rank\n\nMemory keeps winning branches, flags a weak ResNet50. **91.9%** of the gain from Improve + Crossover. Best baseline: 0.2963, no medal.\n\n06 · RL Training Curves · Frontis-MA1-35B\n\nThe 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.\n\n221\n\nRL policy updates logged\n\n0.180 → 0.400\n\nRollout reward · first to final step\n\n0.404\n\nPeak validation base reward\n\n31 /176\n\nPeak validation medal count · out of 176 tasks\n\n**Reward / medals**\n\nstitched training step\n\n07 · Meta-Ability & Transfer\n\nNatureBench Lite, controlled: **hold the model or the framework fixed — both transfer.**\n\nModel transfer · framework fixed · Match-SOTA\n\nSame framework, **Frontis-MA1** swapped in.\n\nFramework transfer · model fixed · Match-SOTA\n\nSame model, **OpenMLE-Evo** search swapped in.\n\n5 / 5\n\nEvery modality group improves. No Medal Rate drops.\n\n+2 +4 +1 +4 +3\n\n14 new medals, spread across all five modalities.\n\n24.2 → 34.8\n\nTest medal rate with upper-tail reward shaping (×4.0).\n\n08 · Open-Source Release\n\nR-01\n\n30B & 35B post-trained checkpoints on HuggingFace.\n\nR-02\n\nAudited task inventory + category-specific release artifacts.\n\nR-03\n\nScheduler + CPU/GPU Docker workers; control / execution / storage separated.\n\nR-04\n\nExecution-grounded SFT data construction + asynchronous RL configs.\n\nR-05\n\nExperience cards, 3-factor selection, operator-conditioned context.\n\nR-06\n\nFull MLE-Bench Lite / NatureBench Lite evaluation & reproduction scripts.\n\n```\n@misc{yang2026frontisma1trainingai4aimodel,\n  title={Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering},\n  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},\n  year={2026},\n  eprint={2607.28568},\n  archivePrefix={arXiv},\n  primaryClass={cs.CL},\n  url={https://arxiv.org/abs/2607.28568},\n}\n```\n\n", "url": "https://wpnews.pro/news/frontis-ma1-open-35b-model-toward-recursive-self-improvement", "canonical_source": "https://frontisai.github.io/OpenRSI/", "published_at": "2026-08-03 00:06:08+00:00", "updated_at": "2026-08-03 00:52:45.822204+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research", "ai-products", "ai-infrastructure"], "entities": ["Frontis AI", "Frontis-MA1", "OpenMLE", "OpenMLE-Gym", "OpenMLE-ERL", "OpenMLE-Evo", "GPT-5.5", "Codex"], "alternates": {"html": "https://wpnews.pro/news/frontis-ma1-open-35b-model-toward-recursive-self-improvement", "markdown": "https://wpnews.pro/news/frontis-ma1-open-35b-model-toward-recursive-self-improvement.md", "text": "https://wpnews.pro/news/frontis-ma1-open-35b-model-toward-recursive-self-improvement.txt", "jsonld": "https://wpnews.pro/news/frontis-ma1-open-35b-model-toward-recursive-self-improvement.jsonld"}}