OpenRSI names its operator learning stack OpenMLE-RL and ships the directory as OpenMLE-ERL

The whole first release is committed to one executable domain

OpenRSI is described as an open initiative for turning AI improving AI into an engineering problem, and it is organised as three connected routes: AI4AI foundation models that internalise reusable research processes, world models and research taste that decide where extra compute is worth spending, and open tasks, environments and evaluations that turn research into training and search pipelines. The loop is described plainly: search produces experience, experience enters training, and trained models return to search and evaluation. The scope limit is stated just as plainly. The mechanism ladder runs from Evolution through Self-Evolution and Meta-Evolution to RSI, and the page says OpenRSI begins at Meta-Evolution, training the improver itself in bounded executable domains, without claiming that general recursive self-improvement has been solved.

Four operators are trained twice, then composed into one search loop

The design decision worth reading is that post-training and inference are aligned to the same action space. Four atomic program-evolution operators are named: Draft, Improve, Debug and Crossover. Those same four are learned from execution-grounded data through supervised fine-tuning and reinforcement learning, on data deduplicated against all evaluation benchmarks, and then composed into long-horizon search. Coupling learning and evolution in a single loop is the stated contribution, and the deduplication step is what makes the later transfer numbers mean something, since the operators never saw a benchmark they are later scored on. The stack around it is named as OpenMLE-Gym for verifiable task environments with execution feedback, and OpenMLE-Evo for long-horizon search.

The headline number is measured on a 4090 capped at 12 GB

The reported result is specific enough to check against. On MLE-Bench Lite, under a 12-hour per-task budget on one RTX 4090 capped at 12 GB of VRAM, Medal Average rises from 39.39% for the base model to 60.61% with OpenMLE-Evo, and to 71.21% with OpenMLE-Evo-Max, which adds benchmark-independent experience priors and asynchronous search. The abstract then says this exceeds GPT-5.5 plus Codex and approaches GPT-5.6 Sol and the 2.8T Kimi K3, without stating what budget or hardware those runs were given. That gap is the part to hold onto: the two halves of the claim, the measurement and the comparison, are documented very differently.

The held-out benchmark swaps one half of the system at a time

NatureBench asks whether coding agents can match the published state of the art of Nature-family papers, and OpenRSI holds it out as a transfer benchmark. The ablation is the useful part, because it moves one component while fixing the other. With the framework fixed and the trained model swapped in, Match-SOTA goes from 50% to 70%. With the model fixed and OpenMLE-Evo swapped in, it goes from 20% to 50%. So neither the weights nor the search loop account for the result alone, and the two contribute different amounts depending on which is held. The 50% to 70% swing is the model, the 20% to 50% swing is the framework, and the paper is explicit that both components transfer.

The operator learning directory is spelled ERL and nothing expands it

The tree at the root holds three project directories: OpenMLE-ERL, OpenMLE-Evo and OpenMLE-Gym. The paper's abstract names the same three components as OpenMLE-Gym, OpenMLE-RL and OpenMLE-Evo, so the directory that should contain operator learning is spelled ERL, and no part of the page says what those three letters stand for. A second naming gap sits in the sandbox, which is not a top-level directory at all but is reached through a README inside OpenMLE-Gym. That sandbox is the self-hosted distributed code-execution and automatic-evaluation backend for OpenMLE-Evo and OpenMLE-RL, with CPU and GPU job scheduling and optional multi-controller routing. Anyone writing automation against this repository has to resolve ERL from the source, not from the page.

No licence is named, and the root still carries LICENSE and NOTICE

The repository metadata does not report a licence, and the page says nothing about terms anywhere. The tree does contain a LICENSE file and a NOTICE file at the root, so the terms are present in the repository and simply not restated on the reading surface. That mismatch matters more here than it would in an ordinary library, because the first release is a model plus a stack plus datasets, all of which are also published outside the repository on Hugging Face. Nothing in the page says whether the weights, the GGUF derivatives and the datasets carry the same terms as the code, and this page does not guess on its behalf. Read both files before redistributing anything from the release.

The weights, the data and the paper all live outside the repository

Nothing in the tree is the artefact you would install. Frontis-MA1 is published on Hugging Face at 35B and at 30B, with GGUF derivatives collected alongside. OpenMLE Tasks and OpenMLE SFT Traces are published there as datasets, and the paper is on arXiv as 2607.28568. The repository has no GitHub releases, so the news list is the only release record, running from NatureBench on 2026-06-23 through the RSI survey on 2026-06-25, the first release on 2026-07-31, the OpenMLE Sandbox on 2026-08-09, and an EMNLP 2026 Main Track acceptance for EEMA on 2026-08-25. The last push to the default branch is 2026-09-17, after every dated entry, so the news list has already fallen behind the code.

Editorial conclusion

OpenRSI is worth reading if you work on execution-grounded agent training and want one loop that couples post-training with search rather than two separate projects, and the OpenMLE stack plus released weights make that loop something you can point at. Do not adopt it as a general recursive self-improvement result, because the page itself limits the claim to bounded, executable domains. Before building on it, read the LICENSE and NOTICE files at the root since the page names no terms, expect the operator-learning code to be under an ERL spelling you will have to map yourself, and treat the model comparisons in the abstract as unverified until you rerun MLE-Bench Lite under the same 12-hour budget.

Frequently asked questions

Official sources

1. FrontisAI/OpenRSI on GitHub
2. Issues
3. Project website
4. README