Frontis-MA1 open-sources 35B recursive self-improvement ML agent

A group of researchers just posted an open recursive self-improvement setup for machine learning engineering, and the specifics are what make it worth reading rather than skimming. Frontis-MA1 is a 35 billion parameter agent, released with weights and the surrounding stack, that its authors say lifts the Medal Average on MLE-Bench Lite from 39.39% to 60.61% over the base model, and up to 71.21% with a fuller search configuration. The paper is on arXiv, with code and weights at FrontisAI/OpenRSI on GitHub.

The interesting bit is the pitch, not just the score. Recursive self-improvement has mostly lived in two places, theory papers and closed labs where you cannot see the training loop. This release goes the other way. The authors put out OpenMLE-Gym for verifiable task environments, OpenMLE-RL for operator learning, and OpenMLE-Evo for long-horizon search, and align post-training and inference around four named operators, Draft, Improve, Debug, Crossover, that get trained end-to-end and then composed into a single loop. If those pieces work as described, someone outside the lab can, in principle, re-run the whole thing.

The compute framing is also unusually modest for a story like this. Each MLE-Bench Lite task ran under a 12-hour budget on a single RTX 4090 capped at 12 GB of VRAM, which is a lot closer to what an academic group actually has than the multi-node setups these numbers usually come with. On the held-out NatureBench Lite, the paper reports that swapping in the trained model raises Match-SOTA from 50% to 70%, and swapping in the search framework raises it from 20% to 50%.

All of this is self-reported, comes from a single paper, and is measured on benchmarks the same team scoped. The claim that Frontis-MA1 exceeds GPT-5.5 + Codex and approaches GPT-5.6 Sol and the 2.8T Kimi K3 is theirs, not an independent one, and the write-up does not detail the post-training compute bill or the license attached to the weights. Until an outside group reproduces this, the numbers should be read as author claims rather than verified results.

Even with those caveats, an open, downloadable RSI stack that runs on hobbyist-grade hardware is the kind of release that changes what a small research group can attempt this quarter. If the numbers survive replication, hosted MLE-copilot vendors will have to explain what their offering does that a locally-run 35B agent does not.