A Chinese research group released Frontis-MA1, a 35-billion-parameter open model post-trained specifically to improve machine-learning code through repeated execution and evolutionary search, along with the entire stack used to build it. Under the same search framework, the trained model raises its base model's medal average on MLE-Bench Lite from 39.39% to 60.61%, and reaches 71.21% with a stronger search configuration - all on one RTX 4090 with its memory capped at 12GB and a 12-hour budget per task.
Key facts
- Medal average on MLE-Bench Lite: 39.39% for the base model, 60.61% for Frontis-MA1 under the same search, 71.21% with the enhanced search mode.
- Constraints: 12-hour budget per task on a single RTX 4090 capped at 12GB of video memory.
- Released July 30, 2026, with model weights and the full OpenMLE stack public.
- Primary source: Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering; code at FrontisAI/OpenRSI; weights at FrontisAI/Frontis-MA1-35B.
The phrase attached to this work - recursive self-improvement - carries more baggage than the work does, so it is worth being precise about what actually happens.
The system trains a model on four specific actions it can take on a program: Draft one, Improve one, Debug one, and Crossover two of them into a hybrid. Those four operators are learned from execution feedback - the code is actually run in a sandbox, and whether it worked and how well it scored becomes the training signal. Then, at test time, the same four operators drive a long-horizon search: keep a population of candidate programs, run them, record what happened, breed and mutate the survivors. Training and inference are aligned around the same small vocabulary of moves.
The analogy is a plant breeder rather than a philosopher. The system does not reason its way to a better model; it grows a few hundred, kills the ones that fail, crosses the ones that succeed, and does that faster and more sensibly than random search because it has been trained on what previous generations of that process looked like.
Two results make the claim harder to dismiss as benchmark-fitting. First, the gain decomposes cleanly: on a held-out benchmark called NatureBench Lite, holding the framework fixed and swapping in the trained model raises the score from 50% to 70%, while holding the model fixed and swapping in the search framework raises it from 20% to 50%. The model and the scaffolding each contribute, and each transfers to tasks it was not tuned on. Second, everything is released - weights, gym, training code, evolution loop - which is what makes independent replication possible at all.
The authors themselves draw the line the headlines will not. The system, they write, is "an open full-stack system for RSI research" - research toward it, using machine-learning engineering as "a concrete, executable testbed." Nothing in the loop lets the deployed system rewrite its own weights, its own training recipe, or its own search policy. Candidate solutions evolve. The thing doing the evolving does not.
It is also worth being precise about what "recursive" refers to here, because the word is doing a lot of work. The recursion runs across time rather than within a run: search trajectories generated by the system become training data for the operators, and the trained operators then drive future searches. That is a genuine feedback loop and a legitimate research contribution. It is not a loop in which a running system decides to change its own objectives, curriculum, search policy or weights - which is what the phrase implies to most readers, and what the authors carefully do not claim.
There is a second caveat that applies to the number rather than the claim. The official MLE-bench repository currently pauses new leaderboard submissions while its maintainers work on fairness and comparability, and lists known defects in tasks that overlap the Lite split - including one with a discoverable held-out source and one with a field that leaks the outcome. That does not invalidate the base-versus-trained comparison, since both sides ran on the same tasks. It does mean the absolute medal percentages, and any "beats model X" framing built on them, should stay provisional until there is a repaired and versioned run. The controlled gain is the durable part of this result; the leaderboard-shaped number around it is not.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)