RSI-Anything is a method study of recursive self-improvement (RSI) that takes JEPA-Anything as its cross-domain predictive foundation.
Recursive self-improvement aims to turn what a system learns in one round of interaction into lasting improvements it still uses when it acts and learns in the next. Two things get in the way. External feedback is expensive: running tests or experiments costs something, so only a few candidates can be checked for real. And once the policy changes, so do the candidates it proposes, and a world model that predicted well before may no longer do so. Improving the policy and the world model separately does not guarantee that the pair works better together.
RSI-Anything uses world-model predictions to choose which candidates go out for external evaluation and which experience is used for learning. When real feedback comes back it improves the policy or the agent workflow, and the world model is adapted only when its predictions are judged unreliable and there is enough data and budget to do so. The two can update on their own schedules. Each new version first has to pass its own check, and the assembled pair is then compared with the current one under the same conditions; it is kept only if it meets criteria fixed in advance. The pair that is kept shapes which candidates the next round proposes and what experience it gathers, which is what makes the improvement recursive.
The paper evaluates the framework on general computational tasks (code repair and GUI interaction), laboratory planning and experimental science (reaction optimization), and presents an end-to-end immune-science application that links agent-workflow evolution, construction of an immune world model, analysis with self-correction, and prospective therapeutic hypothesis generation.
In code repair the coder is fixed to ReasonFlux-Coder-7B, which proposes four candidate revisions per round, of which one is executed. Across 200 CodeContests tasks and three seeds, world-model-guided selection raises the final hidden-test pass rate from 26.3% with a static verifier to 29.2%, and adapting the world model with execution feedback brings it to 31.2%, with about the same number of online test calls per task. In GUI interaction, success on the MiniWoB++ checkbox task rises from 51% to 99% while average actions per episode fall by 39.4%; on OSWorld, task success improves by 18.4% relative to the initial agent and by 5.5% relative to RLAnything.
In laboratory planning, five successive versions of one 4B policy, produced by a single self-improvement process, raise the LabHorizon F200 Level-2 score step by step from 21.80% to 46.01%. In reaction optimization, both Suzuki–Miyaura screens reach a higher best yield than Bayesian optimization after the same number of experiments; in the flow LC–MS screen the mean best observed yield after 40 experiments is 96.15%, against 82.16% for Bayesian optimization. The immune-science application ends with a prospective therapeutic hypothesis, SIRPα inhibition combined with IL-36γ-axis modulation, which awaits experimental testing.
Each evaluation tests a different part of the framework. Whether the policy and the world model keep improving each other over many rounds within one task still needs joint multi-round evaluation; the laboratory results assess protocol-aligned planning, and real execution has to be checked on the relevant equipment and workflows; the chemistry results cover the two reported screens and their experiment budgets. These are observations on selected domains and tasks.
The technical report, the GitHub repository and the models on Hugging Face are public.