AI research firm PhAI Labs has released RSI-Anything, a framework aimed at helping AI agents accumulate experience from real tasks and improve how they work over time. The setup is intended for code repair, computer-use tasks, and scientific research.
At the center of the framework is a "world model" that predicts what different actions are likely to produce. An agent first proposes several candidate plans. The world model then estimates which one is more likely to succeed, and the system executes that choice. Afterward, it compares the prediction with the actual result and adjusts the agent’s operating strategy. If the prediction itself was off, the world model is updated as well. A new version is kept only after it passes testing and is then used in the next round of tasks.
Code repair results
In the bug-fixing example described by the team, the coding model generates four patch candidates in one pass. RSI-Anything predicts which of those changes is most likely to pass testing, runs only that one, and then uses the test outcome to improve later decisions. The idea is to spend limited test attempts on the option with the best chance of working.
Across 200 CodeContests problems, the coding model, ReasonFlux-Coder-7B, did not change during the experiment, and the number of test runs stayed roughly the same. With a traditional static verifier, the bug-fix pass rate was 26.3%. After adding world-model-based filtering, the rate rose to 29.2%. When the world model was then updated using real feedback, the pass rate reached 31.2%.
Applied to computer use and scientific experiments
The team also used the method in computer-use tasks and scientific experiments. In a lab planning test, a 4B model was updated through five consecutive versions, and its planning score increased from 21.80% to 46.01%.
PhAI Labs said the project code has already been made public. At the same time, different tasks still rely on their own separate agents and world models. The paper also says more experiments are needed to verify whether execution policies and world models can continue improving together over the long term within the same task.
