PhAI Labs releases RSI-Anything technical report, exploring cross-domain recursive self-improvement with world models

Recursive self-improvement (RSI) is becoming an important direction in AI research, but for AI to move from digital tasks like code and math to computer environments requiring continuous interaction and even the physical and scientific world, explicit test results and reward signals are not enough. State changes and action consequences in different environments are more complex, and real feedback may be more expensive, sparser, or even difficult to obtain directly.

PhAI Labs released the technical report "RSI-Anything: Unlocking Recursive Self-Improvement with World Models" on October 12, proposing exactly such a path: first build predictive models for different worlds, anticipate consequences before actually acting, then use real feedback to improve itself. This research continues the earlier JEPA-Anything technical line, which explored modeling the change patterns of different worlds—vision, molecules, physics, life, and others—through general predictive learning principles.

What must first be clarified is that "Anything" here does not mean a single world model can already understand all things, or that a single Agent can already self-evolve in any environment. According to the article, what it truly seeks to unify is the underlying mechanism from prediction and interaction to self-improvement across different worlds.

Using world models to assist decision-making is not a new concept; from Model-Based RL to Dreamer and MuZero, AI can already plan and learn with the help of environment prediction. What RSI-Anything focuses on is a further question: how to make prediction, action, real feedback, and the updates of both the policy and the world model itself form a recursive process reusable across domains.

Its core pipeline can be summarized as State → Action → Prediction → Feedback → Improvement. The Agent first generates candidate actions, the world model predicts the possible consequences of different choices, helping the system judge which options are more worth executing; after real execution, the feedback is not only used to solve the current task but can also continue to support policy learning, workflow revision, or world model adaptation.

There is an easily overlooked coupling problem here: when the Agent's policy is updated, it may explore states never encountered before, and the originally accurate world model may no longer reliably predict new actions; conversely, after the world model is updated, it may also change the Agent's judgment of candidate actions. To this end, RSI-Anything allows the Policy and the World Model to update asynchronously and introduces component evaluation and compatibility checks—only policy and world model combinations that pass the checks are retained as the starting point for the next round of improvement.

However, the report does not explain the specific criteria and pass rate of the compatibility checks. This means the key mechanism of "which updates are accepted and why" remains a black box for now.