Frontis-MA1：清华 + Horizon Research 开源 AI 改进 AI 的全套配方，35B 单卡超 GPT-5.5

by Mycelium Protocol

---

「AI 改进 AI」一直是 AI 研究里最诱人、也最模糊的方向——大量论文声称朝着这个目标努力，但很少有工作把它做成可执行、可测量、可复现的工程问题。

OpenRSI（Frontis.AI + 清华大学 + Horizon Research）是一个直接冲着这件事去的开源框架。首个发布是 Frontis-MA1：一个专门为机器学习工程（MLE）后训练的 AI4AI 模型（35B/30B），配套 OpenMLE 全栈，在单张 RTX 4090、12 小时预算内，把 MLE-Bench Lite Medal Average 从 39.39% 跑到了 71.21%——超越 GPT-5.5 + Codex，接近 GPT-5.6 Sol 和 2.8T 参数的 Kimi K3。

论文：arXiv 2607.28568

---

四个字概括：训练「改进者」

**递归自改进（RSI）**的核心不是让模型改进自身，而是让模型成为「能改进 AI 开发过程」的 Agent。Frontis 的路线是把这件事分层：

Evolution → Self-Evolution → Meta-Evolution（当前）→ RSI（目标）

当前 OpenRSI 做到的是 Meta-Evolution：在有界可执行域（MLE）里，训练能改进 AI 程序的 Agent，让每一代 AI R&D 更快、更高效、更可归因。

---

四个原子算子：Draft / Improve / Debug / Crossover

Frontis-MA1 的行动空间只有四个可训练的原子算子，同时用于训练和推理：

算子	功能
Draft	从零生成程序
Improve	利用执行反馈精化父程序
Debug	修复失败程序
Crossover	重组两个父程序

这四个算子贯穿整个系统——SFT 数据用它们标注，RL 训练用它们采样，长视野搜索（OpenMLE-Evo）用它们组合。同一套算子在训练和推理里语义一致，不存在”训练时学一套，推理时用另一套”的脱节。

---

OpenMLE 全栈

OpenRSI 的基础设施由四个模块构成：

OpenMLE-Gym：任务环境

构建、描述、执行和质量检查可验证的 MLE 任务包。每个任务有独立的执行环境和程序化验证器，执行反馈是训练信号的来源。

OpenMLE-RL：算子学习

两阶段训练：

1. 执行反馈 SFT：用 pass@k 成功的轨迹做监督冷启动
2. 在线 RL：以执行结果为奖励信号，在线强化学习

开源数据集：

• OpenMLE-SFT-Traces：SFT 阶段的监督轨迹
• OpenMLE-Tasks：经过审计的可执行任务包

Frontis-MA1：AI4AI 模型

在 OpenMLE 上后训练的元演化 Agent：

• 35B（BF16 + GGUF）：旗舰模型
• 30B（BF16 + GGUF）：备选尺寸

OpenMLE-Evo：长视野搜索

把四个算子组合成迭代搜索：

• 标准 Evo：单 GPU、顺序演化
• Evo-Max：多 GPU 异步搜索 + 独立于 benchmark 的经验先验

---

结果：严格控制对比

这个项目的结果报告方式值得专门说一下——他们明确区分了模型收益和搜索系统收益：

MLE-Bench Lite（12h/任务，1x RTX 4090 12GB）

配置	Medal Average
基础模型（base）	39.39%
+ Frontis-MA1 后训练，Evo 固定	60.61%
+ OpenMLE-Evo-Max（模型+系统全开）	71.21%

71.21% 的结果包含了搜索系统改进（异步多 GPU + 经验先验），不是纯模型分数，README 里明确标注了这一点——这是少见的诚实。

NatureBench Lite（未见评测，迁移验证）

配置	Match-SOTA
基础模型 + 固定 adapter	50%
Frontis-MA1 + 固定 adapter（模型改善，框架不变）	70%
基础模型 + OpenMLE-Evo（框架改善，模型不变）	50%
基础模型 + 固定 adapter（对照）	20%

模型贡献和框架贡献都在 held-out benchmark 上分别验证——这是做 RSI 研究必须要有的实验设计。

---

快速上手

git clone https://github.com/FrontisAI/OpenRSI.git
cd OpenRSI

按目标选入口：

目标	入口
构建或评测可执行任务包	OpenMLE-Gym/README.md
生成 SFT 数据并启动监督训练	OpenMLE-ERL/SFT/README.md
配置并启动执行反馈 RL	OpenMLE-ERL/RL/README.md
运行 Evo 或 benchmark 适配器	OpenMLE-Evo/README.md

模型权重直接从 HuggingFace 拉：FrontisAI/Frontis-MA1-35B（BF16）或 GGUF。

---

为什么值得关注

可执行性是 OpenRSI 的核心主张。AI4AI 领域不缺概念框架，缺的是能在真实机器上跑通、有程序化验证器、结果可复现的端到端系统。OpenRSI 把整条链——任务环境、训练数据生成、SFT、RL、长视野搜索——都开源出来，研究者可以在 OpenMLE 任务上复现，也可以贡献新任务、新算子、新搜索策略。

「改进者自身也是可训练的」这个定位和 TMax（AllenAI 终端 Agent）、EvoScientist 等方向形成互补：TMax 聚焦终端命令执行，OpenRSI 聚焦 ML 程序演化；前者 outcome-only RL，后者 SFT warm-start + 在线 RL + 演化搜索。

CC BY-NC 4.0，非商业可用。188 stars（2026-07-31 才发布），还在快速增长。

仓库：github.com/FrontisAI/OpenRSI · 论文：arXiv 2607.28568 · 模型：HuggingFace FrontisAI

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Frontis-MA1: Tsinghua + Horizon Research Open-Source the Full Recipe for AI That Improves AI — 35B Beats GPT-5.5 on a Single GPU

by Mycelium Protocol

“AI improving AI” has always been one of the most tempting — and most vague — directions in AI research. Plenty of papers claim to work toward it, but few turn it into an executable, measurable, reproducible engineering problem.

OpenRSI (Frontis.AI + Tsinghua University + Horizon Research) is an open-source framework that tackles this directly. Its first release is Frontis-MA1: an AI4AI model (35B/30B) post-trained specifically for machine learning engineering (MLE), with the full OpenMLE stack. On a single RTX 4090 with a 12-hour budget, it pushes MLE-Bench Lite Medal Average from 39.39% to 71.21% — surpassing GPT-5.5 + Codex and approaching GPT-5.6 Sol and 2.8T-parameter Kimi K3.

Paper: arXiv 2607.28568

The Core Idea: Training the Improver

Recursive self-improvement (RSI) isn’t about a model rewriting itself — it’s about making a model that can improve the AI development process. Frontis’s approach is layered:

Evolution → Self-Evolution → Meta-Evolution (current) → RSI (goal)

OpenRSI’s current stage is Meta-Evolution: training an agent that can improve AI programs in a bounded, executable domain (MLE), making each generation of AI R&D faster, more efficient, and more attributable.

Four Atomic Operators: Draft / Improve / Debug / Crossover

Frontis-MA1’s action space consists of four trainable atomic operators, used consistently across both training and inference:

Operator	Function
Draft	Generate a program from scratch
Improve	Refine a parent program using execution feedback
Debug	Repair a failing program
Crossover	Recombine two parent programs

These four operators run through the entire system — SFT data is labeled with them, RL trains over them, long-horizon search (OpenMLE-Evo) composes them. The same operator semantics at training time and inference time, no representation gap.

The OpenMLE Stack

OpenMLE-Gym builds, describes, executes, and quality-checks verifiable MLE task packages. Each task has its own isolated execution environment and a programmatic verifier — execution feedback is the training signal.

OpenMLE-RL trains operators in two stages:

1. Execution-grounded SFT: successful pass@k trajectories as supervised cold-start
2. Online RL: execution results as reward signal, online reinforcement learning

Open datasets: OpenMLE-SFT-Traces (supervised trajectories) and OpenMLE-Tasks (audited task artifacts).

Frontis-MA1 is the post-trained meta-evolution agent — 35B and 30B in BF16 with GGUF derivatives.

OpenMLE-Evo composes the operators into iterative search:

• Standard Evo: single GPU, sequential evolution
• Evo-Max: multi-GPU async search + benchmark-independent experience priors

Results: Controlled Comparisons

The results section is worth highlighting for its methodology — they explicitly separate model gain from search-system gain:

MLE-Bench Lite (12h per task, 1× RTX 4090 12GB):

Configuration	Medal Average
Base model	39.39%
Frontis-MA1 post-trained, Evo fixed	60.61%
OpenMLE-Evo-Max (model + system)	71.21%

The 71.21% result includes search-system improvements (async multi-GPU + experience priors) and is clearly not presented as a pure model score — a level of honesty rarely seen.

NatureBench Lite (held-out, transfer validation):

Configuration	Match-SOTA
Base model + fixed adapter	50%
Frontis-MA1 + fixed adapter (model improved, framework fixed)	70%
Base model + OpenMLE-Evo (framework improved, model fixed)	50%

Model contribution and framework contribution are each validated separately on a held-out benchmark — the experimental design RSI research actually requires.

Quick Start

git clone https://github.com/FrontisAI/OpenRSI.git
cd OpenRSI

Goal	Start here
Build or evaluate executable task packages	OpenMLE-Gym/README.md
Generate SFT data and launch supervised training	OpenMLE-ERL/SFT/README.md
Configure and launch execution-grounded RL	OpenMLE-ERL/RL/README.md
Run OpenMLE-Evo or a benchmark adapter	OpenMLE-Evo/README.md

Model weights: FrontisAI/Frontis-MA1-35B (BF16) or GGUF.

Why This Matters

Executability is OpenRSI’s core claim. The AI4AI space isn’t short on conceptual frameworks — what’s rare is an end-to-end system that runs on real hardware, has programmatic verifiers, and produces reproducible results. OpenRSI open-sources the entire chain: task environments, training data generation, SFT, RL, long-horizon search. Researchers can reproduce results on OpenMLE tasks, or contribute new tasks, operators, and search strategies.

“The improver itself is trainable” positions OpenRSI as a complement to projects like TMax (AllenAI’s terminal agent RL) and EvoScientist: TMax focuses on terminal command execution; OpenRSI focuses on ML program evolution. The training approaches differ too — outcome-only RL vs. SFT warm-start + online RL + evolutionary search.

CC BY-NC 4.0, non-commercial use. 188 stars (published 2026-07-31), growing fast.

Repository: github.com/FrontisAI/OpenRSI · Paper: arXiv 2607.28568 · Models: HuggingFace FrontisAI

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.