Frontis.AI / Horizon Research open-source OpenMLE (gym, RL, Evo) and Frontis-MA1-35B, lifting MLE-Bench Lite medal average to 71.21% under a single RTX 4090 budget toward executable RSI research.
Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering
Horizon Research, Frontis.AI · Tsinghua University
Open weights · open gym · open search — the full OpenMLE stack, released
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.
Stack overview
OpenMLE-Gym — a gym, not a dataset: thousands of executable tasks with structured sandbox feedback modes (MLE-Bench excluded from training).
OpenMLE-ERL — execution-grounded SFT + RL with asynchronous rollouts.
OpenMLE-Evo — test-time scaling toward test-time learning with experience cards and operator-conditioned memory.
Frontis-MA1 (30B / 35B) — trained by OpenMLE, driving OpenMLE, evaluated on third-party benchmarks.
Four operators
Draft (generate from scratch), Improve (refine a parent), Debug (repair failing code), and Crossover (recombine two parents) form a unified action space for code evolution, invoked thousands of times per task.
Results snapshot
System
Medal Average (MLE-Bench Lite)
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
71.21
GPT-5.6 Sol / Kimi K3
72.73
Release
Weights, gym, sandbox, training, search, and evaluation harness are released for reproducible AI4AI / RSI research. Paper: arXiv:2607.28568.