Frontis-MA1 and OpenMLE: A Reproducible AI-Self-Improvement Factory on a Single RTX 4090

Source

• Paper: *Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering* (arXiv:2607.28568, submitted 2026-07-30, 61 pages)
• Institutions: Frontis.AI (Horizon Research) and Tsinghua University Collaborative Interactive Intelligence Research Center (Bowen Zhou, Ning Ding teams), with collaborators from Zhejiang University, Shanghai Jiao Tong University, and Georgia Tech.
• Code & weights: github.com/FrontisAI/OpenRSI, HuggingFace FrontisAI/frontis-ma1

One-Sentence Summary

Rather than a philosopher who drafts one solution and hopes it works, Frontis-MA1 acts as a breeder: it sows many seeds in OpenMLE-Gym (5,758 executable ML engineering plots), runs and scores each, breeds the high-yielding parents with novel branches, and replants—turning "AI improving AI" into a reproducible factory runnable on a single consumer GPU. The crucial ingredient is that scoring is objective (hidden tests + metric.py), not the model's self-assessment.

Key Points

System Identity and Architecture

• Frontis-MA1 is positioned as an open, full-stack OpenMLE system plus a 35B meta-evolution agent. The 35B MoE base uses Qwen3.6-35B-A3B with only ~3B active parameters per token, so the "35B on a 4090" claim refers to total parameters, not dense compute.
• A companion 30B variant uses Qwen3-30B-A3B-Thinking-2507.
• Three-layer stack:
1. OpenMLE-Gym: 5,758 quality-gated executable tasks (156 curated anchors + 3,362 Kaggle dataset tasks + 2,240 Kaggle competition tasks), spanning tabular, text, time-series, and image modalities. 2. OpenMLE-ERL: training layer using SFT (26,259 samples from parallel and evolutionary paths) and GSPO reinforcement learning over four operators. 3. OpenMLE-Evo: experience-driven search with structured Experience Cards, three-factor parent selection, and operator-conditioned on-demand memory synthesis.
• Four operators (Draft / Improve / Debug / Crossover) share the same interface between training and inference.

Training Math: From Noisy MLE Feedback to a Learnable Signal

Three transformations address MLE's delayed, noisy, heterogeneous feedback:
• Adaptive reward bounds normalize per-task scores using on-policy historical frontiers:

r = clip( (s − s_min) / (s_max − s_min), 0, 1 )

• Entropic advantage amplifies top candidates within a group:

A_ent,i ≈ exp(β·r_proc,i) / ( (1/(G−1))·Σ_{j≠i} exp(β·r_proc,j) ) − 1

giving roughly a 4× boost to the best candidate's learning signal.

• Three-factor parent utility combines score, progress, and novelty:

U_i = λ_s·s̃_i + λ_Δ·e^{Δ_i} + λ_n·ν_i

with hyperparameters tuned manually.

• Engineering details: GSPO is the RL algorithm; operator sampling is Draft 0.50 / Improve 0.17 / Debug 0.17 / Crossover 0.16; asynchronous rollout cuts average step time from 97.0 to 50.8 minutes.

Long-Horizon Search Innovations

• Operator-conditioned on-demand memory synthesis: rather than stuffing full history into context, the system retrieves ancestors/siblings/failures relevant to the currently invoked operator. Improve looks at related ancestors; Debug searches by error type; Crossover extracts complementary method components plus known conflicts.
• In 66 matched comparisons against AIRA-Evo on the same 35B model, seed, and 12-hour budget, tokens dropped from 129.3M to 75.3M (−41.7%), evaluated nodes fell only 12.4%, and best-update rate per million tokens rose from 1.77 to 3.27 (+84.3%). The gain comes from better node selection, not fewer experiments.

Main Results: MLE-Bench Lite (22 tasks × 3 runs)

• Qwen3.6-35B-A3B base: 39.39% medal average
• Frontis-MA1-35B + OpenMLE-Evo: 60.61% (+21.22pp from training)
• Frontis-MA1-35B + OpenMLE-Evo-Max: 71.21% (+10.6pp from cross-task priors + async tree search)
• Compared with GPT-5.5 + Codex (68.18%), the system approaches GPT-5.6 Sol + Codex (72.73%) and Kimi K3 + Claude Code (72.73%).
• Human Rank: 0.5828 → 0.7647 → 0.8126.
• The 30B variant reproduces the trend: 34.85% → 53.03% → 66.67%.
• Important caveat: 71.21% is a system score (model + search framework), not a single-model score. Standard deviations are large (e.g., ±8.57% for Evo-Max).

Cross-Task Transfer: NatureBench Lite (10 scientific tasks)

A controlled decomposition separates model contribution from framework contribution:
• Fixed framework, swapped trained model: Match-SOTA 50% → 70%
• Fixed model, swapped OpenMLE-Evo: Match-SOTA 20% → 50%
Both transfer independently, suggesting the gains are not overfitted to MLE-Bench-specific tricks.

Mechanism Analyses

• Late-stage operations dominate value: in leaf-classification, later Improve + Crossover contributed 85% of validation gains; in bird-audio, 91.9%. Real value comes from continually recombining complementary branches and upgrading backbones, not from merely producing a runnable program.
• Targeted Crossover case (nomad2018 material-property prediction): AIRA-Evo ran seven Debug attempts on a broken branch; OpenMLE-Evo crossed a physics-feature-strong branch with a .xyz-stable branch, achieving 8.2% lower validation RMSE and 11.3% lower test RMSE.
• Three-factor selection case (right-whale): selection probability for a lower-scoring but high-progress, structurally novel parent rose from 10.47% to 17.09%, yielding a child held-out AUC of 0.99386.

Competitive Landscape

• AIDE: inference-time search only, no model training.
• AIRA / AIRA-Evo: iterative research framework; OpenMLE-Evo's direct baseline.
• Frontis-MA1 / OpenMLE: training and search share the same operator interface; experience flows back into training.
• BAAI AREX: dual-loop autonomous research with 10B active parameters; emphasizes long-horizon state.
• Discovery Loop (Jeff Dean, 2026): a public-benefit company compressing the hypothesis-result cycle.

All three 2026 efforts share the theme "trace-to-capability": experience must be captured, organized, allocated to the right carrier (skill, memory, harness, or parameters), validated, and re-fed into future behavior. Frontis's distinctive contribution is making verification feedback a reusable training signal and releasing the full stack.

Limitations and Cautions

• The system is the first generation of Meta-Evolution, not RSI. What evolves is external ML artifacts (Kaggle solutions); the search strategy, tasks, and evaluators remain fixed. Hand-tuned hyperparameters (λ, β, adaptive bounds) are still heavy.
• MLE-Bench Lite (22 tasks) and NatureBench Lite (10 tasks) are Kaggle-style structured competitions, missing real industrial scenarios such as dirty-data cleaning, domain-knowledge feature engineering, and deployment constraints. Metrics are scalar, missing maintainability and inference efficiency.
• The official MLE-bench repository has paused new submissions and lists known flaws in the Lite split (one task with discoverable hidden source, one with leaked fields). "Beating X" rankings should be treated as provisional until a fixed split is used.
• SFT deduplication granularity is not detailed, and Kaggle data and code are widely circulated online; base-model pretraining contamination cannot be fully ruled out. NatureBench transfer partially mitigates but does not eliminate this concern.
• Compute cost is non-trivial: 12 hours per task on an RTX 4090 plus many rollouts and retries. Suitable for research labs today, not casual developers.
• "Engineers will lose their jobs" is commentator embellishment, not a paper claim. The demonstrated role is competent MLE intern (write, edit, debug, fuse, run, get objectively scored); humans still define tasks, verifiers, budgets, novelty thresholds, and deployment gates.

Reproducibility Resources

• Paper: https://arxiv.org/abs/2607.28568
• Code: https://github.com/FrontisAI/OpenRSI
• Model weights: https://huggingface.co/collections/FrontisAI/frontis-ma1