AIAny
Icon for item

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

Autonomously proposes, modifies, executes, and evaluates ML experiments to study recursive self-improvement in machine learning engineering. Implements an open stack (OpenMLE-Gym, -RL, -Evo) and post-trains Frontis-MA1 (35B) around four evolution operators (Draft, Improve, Debug, Crossover); releases model weights and the full codebase.

Introduction

Why this matters

Recursive self-improvement (RSI) is the next-step question for AI research: can systems not just solve tasks, but improve the process of building better AI? This paper takes RSI from theory to an executable testbed by centering on machine learning engineering (MLE) as a concrete, measurable domain and by training a 35B meta-evolution agent (Frontis-MA1) that both learns operator behaviors and participates in long-horizon evolutionary search.

Key Findings
  • Combining execution-grounded supervised fine-tuning and reinforcement learning on four atomic program-evolution operators (Draft, Improve, Debug, Crossover) produces a reusable agent that meaningfully improves ML workflows.
  • On MLE-Bench Lite with a 12-hour per-task budget on one RTX 4090 (12GB VRAM), Frontis-MA1 plus OpenMLE-Evo raises Medal Average from 39.39% to 60.61%, and to 71.21% with OpenMLE-Evo-Max (asynchronous search and benchmark-independent priors). This outperforms a GPT-5.5+Codex baseline and approaches much larger specialized systems.
  • Components transfer: swapping the trained model into a fixed OpenMLE framework raises Match-SOTA on held-out NatureBench Lite from 50% to 70%; swapping in OpenMLE-Evo with the base model raises it from 20% to 50%.
What Sets It Apart
  • Execution-grounded operator learning: the same low-level operators are trained from real execution feedback (not just text), so model outputs can be directly validated and iterated in the loop.
  • Full-stack, reproducible research platform: OpenMLE bundles verifiable task environments (OpenMLE-Gym), operator-level agent training (OpenMLE-RL), and long-horizon evolutionary search (OpenMLE-Evo), enabling repeatable experiments on RSI in MLE.
  • Practical budgeted evaluation: demonstrates measurable gains under realistic single-GPU constraints (12GB VRAM), giving concrete throughput and effectiveness numbers rather than only synthetic benchmarks.
Who it's for and trade-offs

Great fit if you want a reproducible, experiment-driven platform to study AI4AI and recursive self-improvement in a hands-on MLE setting; if you need an off-the-shelf agent to propose, edit, run, and evaluate ML experiments within budgeted compute. Look elsewhere if you need turnkey production model deployment, low-latency inference for end-users, or if your focus is on domains unrelated to program/code evolution (the stack is specialized for ML experiment execution and evolution). The approach requires substantial engineering (execution sandboxes, experiment orchestration) and the results are shown under specific compute and benchmark constraints, so expect work to adapt it to different hardware or problem domains.

Information

  • Websitearxiv.org
  • AuthorsJunlin Yang, Che Jiang, Yu Fu, Tianwei Luo, Can Ren, Weizhi Wang, Kaikai Zhao, Hongyi Liu, Yuxin Zuo, Yuru Wang
  • Published date2026/07/30

More Items

Provides a portable C++ inference runtime to deploy embodied AI models (vision–language–action and world–action) on heterogeneous robot hardware, enabling latency-first batch-1 closed-loop control. Key features include modular multi-rate layers, fused low-latency inference, and extensible head/IO plugins.

Standardizes representation-level evaluation for tabular encoders by exporting row-, column-, and table-level embeddings and probing them with shared lightweight heads across three suites (TRL-CTbench, TRL-Rbench, TRL-DLTE). Supplies curated benchmark assets and task rewrites (50 OpenML tables, 123 targets, a 47,772-table DLTE lake) to enable fair cross-paradigm comparison.

Treats the interface between an LM agent and a computer as a design variable. A custom agent-computer interface (ACI) with concise file-edit, repo-navigation, and test commands plus compact feedback reaches 12.5% pass@1 on SWE-bench, 87.7% on HumanEvalFix.