[research] 4 atomic operators let a 35B agent beat GPT-5.5 on ML engineering

[github-actions[bot]]

🔬 The Finding

Researchers introduced Frontis-MA1, a 35B meta-evolution agent for machine learning engineering that achieves recursive self-improvement by decomposing all actions into just four atomic program-evolution operators: Draft, Improve, Debug, and Crossover. Trained via execution-grounded SFT + RL and composed into long-horizon search, it raised Medal Average on MLE-Bench Lite from 39% → 71%, surpassing GPT-5.5 + Codex — on a single RTX 4090 with 12 GB VRAM.

⚙️ What It Means for Agentic Workflows

Design your agent's action space as atomic, composable operators rather than monolithic task prompts. Draft → Improve → Debug → Crossover mirrors what effective CI/automation loops already do; making these explicit lets you train or prompt-engineer each operator independently and then compose them into search. Execution feedback is the key loop — every operator call gets grounded in real runtime output, which is directly achievable in GitHub Actions or any workflow with shell access.

🔗 Source

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering — July 30, 2026

Generated by Daily Agentic AI Research Digest · 55.6 AIC · ⌖ 11.2 AIC · ⊞ 24.1K · ◷

• expires on Aug 8, 2026, 10:31 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-08-08T10:31:46.366Z.

Closed by Workflow