RSI-Anything enables AI to improve on the fly, fixing more bugs without modifying the programming model.

Beating AI Express: AI research institute PhAI Labs has launched RSI-Anything, a framework enabling AI agents to continuously accumulate experience from real tasks and improve their performance. The framework supports use cases including code debugging, computer operations, and scientific research. Its core design equips agents with a "world model" that predicts the outcomes of different actions. The agent first proposes multiple solutions; the world model evaluates which is most likely to succeed, after which the agent selects that solution for execution. The system then compares predictions with actual results to refine the agent’s operational strategy. If predictions are inaccurate, the world model is also updated. Only after the new version passes tests is it retained for subsequent tasks. Take bug fixing as an example: the code-generating model produces four modification proposals at once. RSI-Anything first predicts which is most likely to pass tests, selects only one to run, then uses the test result to improve future judgments. This approach allocates limited test attempts to more promising solutions. In 200 CodeContests problems, the code-writing model ReasonFlux-Coder-7B remained unchanged throughout, with roughly the same number of test attempts. Using a traditional static verifier, the bug-fixing pass rate stood at 26.3%. After introducing the world model for screening, the rate rose to 29.2%; further updating the world model with real feedback pushed it to 31.2%. The team also applied this method to computer operations and scientific experiments. In a lab planning test, a 4B model was updated five consecutive times, with its planning score rising from 21.80% to 46.01%. The project code is now open-source. Currently, separate agents and world models are used for different tasks. The paper acknowledges that whether the execution strategy and world model can improve together long-term on the same task requires additional experimental verification.