Frontis-MA1-35B: Open-Source 35B AI-for-AI Model Advances Recursive Self-Improvement

Background and Motivation

In early August 2026, Google chief scientist Jeff Dean left the company with several core researchers to launch Discovery Loop, a venture aimed at Recursive Self‑Improvement (RSI) – enabling AI systems to run experiments, evaluate results, fix weaknesses, and iterate autonomously. This reflects a broader trend where AI is being used as a researcher rather than merely a research target, a field often called AI‑for‑AI or Automated AI R&D.

Problem Definition: Trace‑to‑Capability

RSI is framed as the "Trace‑to‑Capability" problem: experience must be captured, organized, assigned to the correct update carrier, verified, and then incorporated as skills, memory, harnesses, or model parameters for future behavior.

Methodology – Frontis‑MA1 and OpenMLE

The technical report "Frontis‑MA1: Training an AI4AI Model toward Recursive Self‑Improvement" introduces the Frontis‑MA1‑35B model and the OpenMLE open‑source toolkit. Machine‑learning engineering (MLE) is chosen as the testbed because an agent both builds another AI system and can be evaluated through real execution and hidden tests.

OpenMLE defines four program operations:

Draft – generate a solution from scratch.

Improve – modify an existing runnable solution.

Debug – fix errors based on logs and failures.

Crossover – recombine compatible parts of two parent solutions.

These operations are used both during training and inference, allowing the model to learn actions that can be directly applied in the search system.

Data Construction

Two data streams are collected:

Parallel paths: 17,245 complete Draft programs.

Evolution paths: 9,014 multi‑step evolution steps selected from high‑quality trajectory fragments.

In total, 26,259 supervised‑fine‑tuning (SFT) samples are created. Only steps that are inherited by later programs and pass quality thresholds are kept, filtering out spurious actions.

Training and Reward Design

Reinforcement learning faces heterogeneous metrics (AUC, RMSE, custom scores). OpenMLE‑ERL normalizes scores within each rollout and applies an Entropic Advantage to weight high‑performing candidates, focusing on relative improvement rather than mere execution success.

Experimental Results

On MLE‑Bench Lite (22 tasks, three runs each on an RTX 4090 for 12 hours, 12 GB memory), Frontis‑MA1‑35B with standard OpenMLE‑Evo achieves a Medal Average of 60.61 %. Adding OpenMLE‑Evo‑Max raises this to 71.21 %.

For comparison, GPT‑5.6 Sol and the 2.8 T‑parameter Kimi K3 score 72.73 % in the same group.

Token usage drops from 129.3 M to 75.3 M (‑41.7 %). New best results per million tokens increase from 1.77 to 3.27, and the proportion of improvements from the Improve operation rises from 4.73 % to 9.36 %.

On the same hardware budget, Frontis‑MA1‑35B outperforms the baseline Qwen3.6‑35B‑A3B (39.39 % vs. 60.61 %). Scaling to 30 B parameters shows a similar jump (34.85 % → 53.03 %).

OpenMLE‑Evo‑Max, which adds cross‑task experience priors and asynchronous multi‑GPU search while keeping total sandbox compute constant, reaches 71.21 %, 3.03 percentage points above GPT‑5.5 + Codex and within 1.52 points of GPT‑5.6 Sol and Kimi K3. Human Rank improves from 0.5828 (base model) to 0.7647 (Frontis‑MA1‑35B) and 0.8126 with Evo‑Max.

Transfer tests on NatureBench Lite (10 scientific tasks) show that swapping only the base model to Frontis‑MA1 raises the number of SOTA‑level tasks from 5 to 7, while swapping only the search framework to OpenMLE‑Evo raises it from 2 to 5, indicating that both model and search contribute.

Analysis of Design Choices

OpenMLE‑Evo reduces prompt length and token count while increasing the discovery of new best results, demonstrating that the performance gain stems from better node selection rather than merely fewer experiments.

In a materials‑property prediction task, OpenMLE‑Evo’s Crossover combines a branch with physical features and another that reliably parses irregular .xyz files, reducing validation RMSE by 8.2 % and hidden‑test RMSE by 11.3 %.

Key Insights and Future Directions

Evaluators must precede self‑improvement; stable measurement of modifications is essential.

The training target should be the modification process itself, not just the final program.

Memory mechanisms need to select relevant history rather than store everything; Experience Cards guide node expansion and code generation.

Meta‑evolution, where the improver itself evolves, is a more accurate description of the current work; full multi‑generation RSI remains an open challenge.

Resources

Paper: https://arxiv.org/pdf/2607.28568

GitHub: https://github.com/FrontisAI/OpenRSI

Hugging Face collection: https://huggingface.co/collections/FrontisAI/frontis-ma1