RSI-Anything: World Models Enable Cross-Domain Recursive Self-Improvement

From Modeling Everything to Letting Everything Drive AI Evolution

To enable AI to continuously self-improve in different environments, a fundamental problem must be solved: how to understand the consequences of actions. In code, running tests provides relatively clear feedback. But in chemistry experiments, changing a reaction condition may require real experiments; in biological systems, an intervention can trigger complex multi-scale changes. If every learning step depends on real trial-and-error, the exploration space is severely limited.

World models offer an alternative: let AI first learn the environment's dynamics, then use predictions to assist action selection. PhAI Labs' earlier JEPA-Anything explored a unified predictive learning mechanism across seven domains — vision, biology, clinical, control, molecular dynamics, physical fields, and weather. Its core method, Orthogonal Predictive Factorization (OPF), decomposes complex states into complementary predictive factors to learn dynamics across systems, demonstrating cross-domain applications from tumor intervention analysis to planetary orbit regularity extraction.

However, predicting the world is only the first step. Even if a world model can predict action consequences, the agent may not know how to choose actions, how to learn, or how to sustain improvement. RSI-Anything attempts to bridge this gap by connecting the world model's predictive capability with the agent's action selection, real feedback, and system updates.

RSI-Anything: How World Models and Agents Form Recursive Improvement

Using world models to assist decision-making is not new — from Model-Based RL to Dreamer and MuZero, AI has leveraged environment predictions for planning and learning. RSI-Anything focuses on how to make prediction, action, real feedback, and updates of both policy and world model constitute a recursive process reusable across domains.

The core loop: State → Action → Prediction → Feedback → Improvement .

Agent generates candidate actions; world model predicts consequences to decide which options are worth executing.

After real execution, feedback not only solves the current task but also supports policy learning, workflow revision, or world model adaptation.

A critical issue: when the policy updates, it may explore previously unseen states; the previously accurate world model may not reliably predict new actions. Conversely, world model updates may change the agent's judgment of candidate actions. The agent changes how it explores the world; the world model must continually adapt to these changes.

RSI-Anything allows asynchronous updates of Policy and World Model, introducing component evaluation and compatibility checks. Not every update is accepted — only policy–world model combinations that pass validation are retained as the starting point for the next improvement round. Thus, the system not only corrects the current answer but attempts to convert experience into capabilities that persistently influence subsequent actions and learning.

Experiments Across Coding, GUI, Laboratory, and Chemistry

(1) Code: Fixing More Bugs Without Upgrading the Coder

In CodeContests code repair experiments, the generator ReasonFlux-Coder-7B was fixed, producing four candidate patches per round and selecting one for execution. All methods shared the same generator and candidate budget. Results: static verifier achieved 26.3% repair success; adding a fixed world model raised it to 29.2%; further adapting the world model with execution feedback reached 31.2% . Average online test invocations stayed around 3.7 per task, showing that improving the world model's ability to judge which actions are worth executing can boost system performance without a stronger generator or more real trials.

(2) GUI Agent: From Code Generation to Computer Interaction

GUI agents must understand visual screen information and interact via mouse/keyboard; each operation changes the interface state, demanding continuous decision-making. On the OSWorld benchmark, RSI-Anything achieved a relative task success rate improvement of ~18.4% , demonstrating potential in complex graphical interaction tasks.

(3) Laboratory: 4B Model, Five Generations, Planning Score Doubles

In LabHorizon laboratory planning tasks, a 4B-parameter policy model underwent five consecutive generations. Planning scores progressed: 21.80% → 38.03% → 41.99% → 44.56% → 46.01% . The final version reached ~2.1× the initial score, slightly surpassing the Actor–Simulator–Selector reference system (45.32%). This shows the system's potential to continuously retain policy updates and gradually improve complex experimental planning. Note: this measures protocol planning quality, not wet-lab execution success.

(4) Chemistry: 40 Experiments, Finding Higher-Yield Reaction Conditions

For chemical reaction optimization, RSI-Anything used a world model to predict outcomes of different catalyst, ligand, solvent combinations, then combined real measurement feedback to select the next experiment round. In two Suzuki–Miyaura reaction screening tasks:

After 20 experiments, best observed yield reached 93.0% , above Bayesian optimization's 87.5%.

After 40 experiments, average best observed yield reached 96.15% , above Bayesian optimization's 82.16%.

For resource-limited scientific research, this means the same experimental budget can yield better exploration results.

Across code testing, GUI interaction, experiment planning, and chemical yield measurement — four tasks with vastly different state spaces, action spaces, and feedback mechanisms — a similar underlying mechanism uses prediction and feedback to improve decision-making and learning.

From Problem-Solving to Exploring the Unknown: RSI Meets Immunology

If code and chemistry focus on improving performance on known tasks, immunology pushes RSI-Anything into a more open research process. The immune system spans cells, tissues, and organism scales; identical interventions can produce drastically different outcomes in different immune contexts. PhAI Labs previously proposed EvoIWM (Evolutionary Immune World Model) to predict multi-scale immune state changes and intervention consequences.

In the RSI-Anything end-to-end demonstration, the team used Agent Genesis to evolve scientific agent workflows, built an immune world model, and applied model-driven analysis, evidence testing, and self-correction to screen potential interventions. The research proposed a candidate combinatorial intervention hypothesis involving SIRPα inhibition and IL-36γ axis modulation , pending experimental validation.

Here, the world model becomes a tool for scientific agents to analyze complex systems, compare intervention strategies, and correct research judgments. This raises a more imaginative direction: when AI learns the dynamics of different worlds, can it not only improve problem-solving but also participate in discovering new problems worth studying?

Anything: Making the Entire World a Training Ground for Intelligence Evolution

JEPA-Anything and RSI-Anything follow a continuous technical line. JEPA-Anything establishes cross-domain predictive learning, letting AI learn how different worlds change. RSI-Anything further connects prediction with real action, external feedback, and system updates, exploring a cross-domain reusable recursive self-improvement framework.

The "Anything" does not imply a single world model understands everything or an agent autonomously evolves in any environment. Rather, it attempts to unify the underlying mechanisms from prediction, interaction, to self-improvement across worlds. Current work validates separate components — prediction selection, model adaptation, policy iteration, feedback-driven search. Long-term, stable co-evolution of policy and world model in a single complex environment remains a future goal.

Future intelligent systems will continually face new tasks, environments, and failure modes. The ability to continuously extract learning signals from these changes may be key to further capability advances. World models provide a potential foundation: making complex environments predictable, using real feedback to correct judgments, and letting verified improvements influence subsequent learning and action. From code to GUI, from digital to physical, to complex scientific exploration. Science is not just a vertical application of RSI; with its complex environments, expensive feedback, and abundant unknowns, it may become a critical testbed for the limits of intelligent self-evolution.

From World Modeling Anything, to Recursive Self-Improvement through Anything.

JEPA-Anything explores how AI can predict different worlds; RSI-Anything further explores how prediction and feedback in those worlds can become resources for continuously improving intelligent systems. From modeling everything, to letting everything drive intelligent evolution.

References

RSI-Anything technical report: https://arxiv.org/abs/2610.13912

RSI-Anything code: https://github.com/Gen-Verse/RSI-Anything

JEPA-Anything paper: https://arxiv.org/abs/2609.20800

JEPA-Anything GitHub: https://github.com/Gen-Verse/JEPA-Anything

PhAI Labs website: https://phai-labs.com/