Gen-Verse/RSI-Anything

What it solves

RSI-Anything addresses the challenge of enabling AI systems to improve their own performance over time across diverse domains. It provides a general framework for recursive self-improvement by allowing a system to predict the outcomes of its actions, interact with an environment, and use the resulting feedback to refine its future decisions.

How it works

Built on the JEPA-Anything predictive foundation, the framework follows a continuous cycle: Propose $\rightarrow$ Predict $\rightarrow$ Select $\rightarrow$ Evaluate $\rightarrow$ Learn $\rightarrow$ Retain.

1. Propose: A policy or workflow suggests potential actions.
2. Predict: A world model anticipates the consequences of those actions.
3. Select: The best action is chosen based on these predictions.
4. Evaluate: The selected action is executed, and external feedback is collected.
5. Learn & Retain: The experience is used to update the system, which then guides the next round of improvement.

Domain adapters allow this shared workflow to be applied to specific tasks by defining the state, action space, and feedback mechanisms for that domain.

Who it’s for

This framework is designed for researchers and developers working on autonomous agents, scientific discovery, and automated software engineering who want to implement self-correcting systems that learn from interaction.

Highlights

• Domain-General: Successfully applied to code repair, GUI interaction, laboratory planning, reaction optimization, and immune science.
• Predictive Guidance: Uses world models to rank candidates before execution, reducing wasted actions.
• Proven Results: Achieves high success rates in code repair (e.g., 46.37% on APPS Introductory) and GUI tasks (99% success on MiniWoB++ checkboxes).
• Scientific Application: Capable of prospective hypothesis generation in immune science and optimizing chemical reaction yields.